What Does My QA Model Know? Devising Controlled Probes Using Expert Knowledge

Automatically answering questions, especially in the open-domain setting where minimal or no contextual knowledge is explicitly provided, requires considerable background knowledge and reasoning abilities. For example, answering the two questions in the top gray box in Figure 1 requires identifying a specific ISA relation (that ‘cooking’ is a type of ‘learned behavior’) as well as recalling a concept definition (that ‘global warming’ is defined as a ‘worldwide increase in temperature’).

Recent success in QA has been driven largely by new benchmarks (Zellers et al., 2018; Talmor et al., 2019b; Bhagavatula et al., 2020; Khot et al., 2020) and advances in model pre-training (Radford et al., 2018; Devlin et al., 2019). This raises a natural question: Do state-of-the-art multiple-choice QA (MCQA) models that excel at standard benchmarks truly possess basic knowledge and reasoning skills expected in these tasks?

Answering this question is challenging because of limited understanding of heavily pre-trained complex models and the way existing MCQA datasets are constructed. We focus on the second aspect, which has two limitations: Large-scale crowdsourcing leaves little systematic control over question semantics or requisite background knowledge (Welbl et al., 2017), while questions from real exams tend to mix multiple challenges in a single dataset, often even in a single question (Clark et al., 2018; Boratko et al., 2018).

To address this challenge, we propose systematically constructing model competence probes by exploiting structured information contained in expert knowledge sources such as knowledge graphs and lexical taxonomies. Importantly, these probes are diagnostic tasks, designed not to impart new knowledge but to assess what models trained on standard QA benchmarks already know; as such, they serve as proxies for the types of questions that a model might encounter in its original task, but involve a single category of knowledge under various controlled conditions and perturbations.

Figure 1 illustrates our methodology. We start with a set of standard MCQA benchmark tasks $D$ and a set of models ℳ trained on $D$⁠. Our goal is to assess how competent these models are relative to a particular knowledge or reasoning skill S (e.g., definitions) that is generally deemed important for performing well on D. To this end, we systematically and automatically generate a set of dataset probesP_S from information available in expert knowledge sources. Each probe is an MCQA rendering of the target information (see examples in Figure 1, yellow box). We then use these probes P_S to ask two empirical questions: (1) How well do models in ℳ already trained on $D$ perform on probing tasks P_S? (2) With additional nudging, can models be re-trained, using only a modest amount of additional data, to perform well on each probing task P_S with minimal performance loss on their original tasks D (thus giving evidence of prior model competence on S)?

While our methodology is general, our experiments focus on probing state-of-the-art MCQA models in the domain of grade-school level science, which is considered particularly challenging with respect to background knowledge and inference (Clark, 2015; Clark et al., 2019; Khot et al., 2020). In addition, existing science benchmarks are known to involve widespread use of definition and taxonomic knowledge (see detailed analysis by Clark et al. [2018], Boratko et al. [2018]), which is also fundamental to deeper reasoning. Accordingly, we use the most widely used lexical ontology WordNet (Miller, 1995) and publicly available dictionaries as sources of expert knowledge to construct our probes, WordNetQA (Section 3.1) and DictionaryQA (Section 3.2).¹ These probes measure competence in various settings including hypernymy, hyponymy, and synonymy detection, as well as word sense disambiguation.

Our exploration is closely related to the recent work of Talmor et al. (2019a). However, a key difference is that they study language models (LMs), for which there is no clear a priori expectation of specific knowledge or reasoning skills. In contrast, we focus on models heavily trained for benchmark QA tasks, where such tasks are known to require certain types of knowledge and reasoning skills. We probe whether such skills are actually learned by QA models, either during LM pre-training or when training for the QA tasks.

Recognizing the need for suitable controls in any synthetic probing methodology (Hewitt and Liang, 2019; Talmor et al., 2019a), we introduce two controls: (a) the probe must be challenging for any model that lacks contextual embeddings, and (b) strong models must have a low inoculation cost—that is, when fine-tuned on a few probing examples, the model should mostly retain its performance on its original task.² This ensures that the probe performance of a model, even when lightly inoculated on probing data, reflects its knowledge as originally trained for the benchmark task, which is precisely what we aim to uncover.

Constructing a wide range of systematic tests is critical for having definitive empirical evidence of model competence on any given phenomenon. Such tests should cover a broad set of concepts and question variations (i.e., systematic adjustments to how the questions are constructed). When assessing ISA reasoning, not only is it important to recognize in the question in Figure 1 that cooking is a learned behavior, but also that cooking is a general type of behavior or, through a few more inferential steps, a type of human activity. Our automatic use of expert knowledge sources allows constructing such high-coverage probes, circumventing pitfalls of solicitation bias and reporting bias.

Our results confirm that transformer-based QA models³ have a remarkable ability to recognize the types of knowledge captured in our probes—even without additional fine-tuning (i.e., in a zero-shot setting). Such models can even outperform strong task-specific non-transformer models trained directly on our probing tasks (e.g., +26% compared to a task-specific LSTM). We also show that the same models can be effectively re-fine-tuned on small samples (even 100 examples) of probe data, and that high performance on the probes tends to correlate with a smaller drop in the model’s performance on the original QA task.

Our comprehensive assessment also reveals important nuances to the positive trend. For example, we find that the best models still perform 2–10% (absolute) below conservative estimates of human performance (Section 3.1.3) on these tasks. Further, the accuracy of even the best QA model degrades substantially on our hyponym probes (by 8–15%) when going from 1-hop hyponym links to 2-hops. The accuracy on the WordNetQA probe drops by 14–44% under our cluster-level analysis (Section 3.1.1), which assesses whether a model knows several facts about each individual concept, rather than only answering correctly isolated questions. This shows that state-of-the-art QA models have much room to improve even in some fundamental building blocks (definitions and taxonomic hierarchies) of more complex forms of reasoning.

We follow recent work on constructing challenge datasets for probing neural models, which has primarily focused on the task of natural language inference (NLI) (Glockner et al., 2018; McCoy et al., 2019; Rozen et al., 2019; Warstadt et al., 2019). Most of this work looks at constructing data through adversarial generation methods, which have also been found useful for creating stronger models (Kang et al., 2018). There has also been work on using synthetic data of the type we consider in this paper (Poliak et al., 2018a; Geiger et al., 2019; Yanaka et al., 2020; Clark et al., 2020). We closely follow the methodology of Richardson et al. (2020), who use hand-constructed linguistic fragments to probe NLI models and study model re-training using a variant of the inoculation by fine-tuning strategy of Liu et al. [2019a]. In contrast, we focus on probing open-domain MCQA models (see Si et al. (2019) for a study on reading comprehension) as well as constructing data from much larger sources of structured knowledge.

Our main study focuses on probing the BERT model and fine-tuning approach of Devlin et al. (2019), and other variants thereof, which are all based on the transformer architecture of Vaswani et al. (2017). There have been recent studies into the types of relational knowledge contained in large-scale knowledge models (Schick and Schütze, 2020; Petroni et al., 2019; Jiang et al., 2019), which also probe models using structured knowledge sources. These studies, however, primarily focus on unearthing the knowledge contained in the underlying language models as is without further training, using simple (single token) cloze-style probing tasks and templates. Most of these results only provide a lower-bound estimate of model performance, since the probing templates being used potentially deviate from what the model has observed during pre-training. In contrast, we focus on understanding the knowledge contained in language models after they have been trained for a QA end-task using benchmark datasets in which such knowledge is expected to be widespread. Further, our evaluation is done before and after these models are fine-tuned on our small samples of target data. This has the advantage of allowing each model to become informed about the format of each probe. We also explore a more complex set of probing templates.

The use of lexical resources such as WordNet to construct datasets has a long history, and has recently appeared in work on adversarial attacks (Jia and Liang, 2017) and general task construction (Pilehvar and Camacho-Collados, 2019). In the area of MCQA, there is related work on constructing questions from tuples (Jauhar et al., 2016; Talmor et al., 2019b), both of which involve standard crowd annotation to elicit question-answer pairs (see also Seyler et al., 2017; Reddy et al., 2017). In contrast to this work, we focus on generating data in an entirely automatic and silver-standard fashion (i.e., in a way that potentially introduces a little noise), which obviates the need for expensive annotation and gives us the flexibility to construct much larger datasets that control a rich set of semantic aspects of the target questions. Following standard practices in MCQA dataset creation (e.g., Khot et al., 2020), however, we perform crowd-sourcing to obtain conservative (in the sense of Nangia and Bowman [2019]) estimates of human performance on our main evaluation sets, to compare against model performance.

Although our probing methodology is amenable to any domain, we focus on probing open-domain QA models in the domain of grade-school level science using a standard suite of benchmark QA datasets (see Table 6). Our choice of this domain is based on the following considerations: It is well-studied qualitatively (Davis, 2016), making it relatively easy to know the types of probes and diagnostic tests to construct using existing expert knowledge. For example, the manual analysis of Mihaylov et al. (2018) found that explicit definitional and ISA knowledge occurred in around 20% and 18%, respectively, of the questions sampled in one benchmark task. Clark et al. (2013) and Boratko et al. (2018) provide similar results involving other benchmarks used in our study.

We also examined MCQA models trained on closely related datasets tailored to commonsense and situational reasoning (Zellers et al., 2018; Talmor et al., 2019b; Bhagavatula et al., 2020; Sap et al., 2019). However, there has been a limited study of the kinds of knowledge needed in this domain, as well as expert knowledge sources for creating corresponding probes. MCQA models trained in this domain exhibit lower performance on our definition and ISA probes.

Our probing methodology starts by constructing challenge datasets (Figure 1, yellow box) from a target set of knowledge resources. Each probing dataset consists of multiple-choice questions that include a questionq and a set of answer choices or candidates {a₁, … a_N}. This section describes in detail the 5 datasets we build (grouped into WordNetQA and DictionaryQA), drawn from two publicly available resources: WordNet (Miller, 1995) and the GNU Collaborative International Dictionary of English (GCIDE).⁴

For convenience, we will describe each source of expert knowledge as a directed, edge-labeled graph G. The nodes of this graph are $V=C∪$$W∪S∪D$⁠, where $C$ is a set of atomic concepts, $W$ a set of words, $S$ a set of sentences, and $D$ a set of definitions (see Table 1 for details for WordNet and GCIDE). Each edge of G is directed from an atomic concept in $C$ to another node in V, and is labeled with a relation, such as hypernym or isa^↑, from a set of relations $R$ (see Table 1).

When defining our probe question templates, it will be useful to view G as a set of (relation, source, target)triples$T⊆R×C×V$⁠. Because of their origin in an expert knowledge source, such triples preserve semantic consistency. For instance, when the relation in a triple is def, the corresponding edge maps a concept in $C$ to a definition in $D$⁠.

We rely on two heuristic functions, defined below for each individual probe: $genQ(τ)$⁠, which generates gold question-answer pairs (q,a) from a set of triples $τ⊆T$ and question templates $Q$⁠, and distr(τ′), which generates distractor answers choices ${a1′,...aN−1′}$ based on another set of triples τ′ (where usually τ ⊂ τ′). For brevity, we will use gen(τ) to denote $genQ(τ)$⁠.

In generating our dataset probes, our general strategy is to build automatic silver-standard training and developments sets, in the latter case at a large scale to facilitate detailed and controlled analysis of model performance. As discussed below, we also provide estimates of human performance on our test sets, and in some cases introduce smaller gold-standard test sets to allow for a direct comparison with model performance.

WordNet is a publicly available English lexical database consisting of around 117k concepts, which are organized into groups of synsets that each contain a gloss (i.e., a definition), a set of representative English words (called lemmas), and, in around 33k synsets, example sentences. In addition, many synsets have ISA links to other synsets that express complex taxonomic relations. Figure 2 shows an example and Table 1 summarizes how we formulate WordNet as a set of triples $T$ of various types. These triples together represent a directed, edge-labeled graph G.

Our main motivation for using WordNet, as opposed to a resource such as ConceptNet (Havasi et al., 2007), is the availability of glosses (⁠$D$⁠) and example sentences (⁠$S$⁠), which allows us to construct natural language questions that contextualize the types of concepts we want to probe. For example, when probing whether a model has knowledge of a concept such as bank (a financial institution), we provide an example sentence he cashed a check at the bank, to help disambiguate the particular sense of bank we are probing. Sentential contexts also provide additional hints to models in cases of rare or infrequent concepts.⁵ Because WordNet is the most authoritative and widely used knowledge resource in NLP, it also has the advantage of having mappings into other knowledge resources (Niles and Pease, 2001; Navigli and Ponzetto, 2010; Tandon et al., 2017), which allows for easily extending our probes to other domains and phenomena.

We build 4 individual datasets based on semantic relations native to WordNet: hypernymy (i.e., generalization or ISA reasoning up a taxonomy, ISA^↑), hyponymy (ISA ^↓), synonymy, and definitions. To generate a set of questions in each case, we use a number of rule templates $Q$ that operate over tuples. A subset of such templates is shown in Table 2 and were designed to mimic naturalistic (i.e., human-authored) questions we observed in our science benchmarks.

For example, suppose we wish to create a question q about the definition of a target concept $c∈C$⁠. We first select a question template from $Q$ that first introduces the concept c and its lemma $l∈W$ in context using the example sentence context $s∈S$⁠, and then asks to identify the corresponding WordNet gloss $d∈D$⁠, which serves as the gold answer a. The same is done for ISA reasoning; each question about a hypernym/hyponym relation between two concepts $c→↑/↓c′∈Ti$ (e.g., dog →^↑/↓animal/terrier) first introduces a context for c and then asks for an answer that identifies c′ (which is also provided with a gloss so as to contain all available context).

For each concept c (an atomic WordNet synset) and probe type (definitions, hypernymy, etc.), we have a wide variety of questions related to c that manipulate (1) the complexity of reasoning that is involved (e.g., the number of inferential hops) and (2) the types of distractors (or distractor perturbations) that are used. We call such sets semantic clusters.

Table 3 shows three examples, capturing ISA reasoning about the following target concepts: trousers, opposing, and poet laureate. Such clusters enable new types of evaluation of the comprehensiveness and consistency of a model’s knowledge of target concepts.

Details of the individual datasets, including average cluster sizes, are summarized in Table 4.

From these sets, we follow Richardson et al. (2020) in allocating a maximum of 3k examples for inoculating the models in the manner described in the next section (i.e., for continuing to train QA models and introduce them to the format of our probes), and reserve the rest for development and testing. In particular, we build large development sets, which are important for performing detailed analysis and cluster-based evaluation.

We report human scores on the individual test sets in WordNetQA (see bottom of Table 7). This is done in two ways.

First, for our test sets generated for definitions and synonyms that cover a large set of disconnected concepts in the WordNet graph and where it is infeasible to annotate individual instances of concepts, we estimate human performance by having crowd-workers on Amazon Mechanical Turk answer a random sample of 500 test questions. Scores are computed by taking the majority vote for each question among 5 annotators. This follows exactly the evaluation protocol used by Nangia and Bowman (2019) and is a conservative estimate in that crowd annotators received virtually no training and no qualification exam before participating in the task.

Second, for our hypernymy and hyponymy test sets, which cover a smaller number of densely connected concepts, we annotated smaller gold-standard test sets that include a sample of around 2,000 random questions that cover a large proportion of the concepts being probed and that have high human performance. To do this, we follow the annotation strategy described above, and greedily apply filtering to remove questions incorrectly answered by human annotators, which follows prior work on building evaluation sets for MCQA (Mihaylov et al., 2018; Talmor et al., 2019b; Khot et al., 2020).

The DictionaryQA dataset is created from the English dictionary GCIDE built largely from the Webster’s Revised Unabridged Dictionary (Webster, 1913), which has previously been used in other NLP studies because of its large size and public availability (Hill et al., 2016). Each dictionary entry consists of a word, its part-of-speech, its definition, and an optional example sentence, as shown for an example in Table 5.

Overall, 33k entries (out of a total of 155k) contain example sentences/usages. As with the WordNet probes, we focus on this subset so as to contextualize each word being probed. Because GCIDE does not have ISA relations or explicit synsets, we take each unique entry to be a distinct sense. Our probe centers around word-sense disambiguation.

To buildQA examples, we use the same generation templates for definitions exemplified in Table 2 for WordNetQA. To construct distractors, we simply take alternative definitions for the target words that represent a different word sense (e.g., the alternative definitions of gift in Table 5), and randomly chosen definitions if needed to create a 5-way multiple choice question. As above, we reserve a maximum of 3k examples for training, and use the same amount for development.

Our initial attempts at building this dataset via standard random splitting resulted in certain systematic biases, revealed by high performance of the choice-only model we used as a control. Among other factors, we found the use of definitions from entries without example sentences as distractors (see again Table 5) to have a surprising correlation with such biases. Filtering such distractors helped improve the quality of this probe.

For assessing human performance, we annotated a smaller gold-standard test set consisting of around 1,100 questions using the crowd-sourcing elicitation setup described in Section 3.1.

Given the probes above, we now can start to answer the empirical questions posed at the beginning. Our main focus is on looking at transformer-based MCQA models trained on science benchmarks in Table 6. We start with our target MCQA models, as well as several control baselines.

Given a dataset $D={(q(d),{a1(d),…,aN(d)})}d∣D∣$ consisting of pairs of questions stems q and answer choices a_i, the goal is to find the correct answer $ai*$ that correctly answers each q. Throughout this paper, we look at 5-way multiple-choice problems (i.e., where each N = 5).

We specifically use BERT-large uncased with whole-word masking, as well as the RoBERTa-large model from Liu et al. (2019b), which is a more robustly trained version of the original BERT model. Our system uses the implementations provided in AllenNLP (Gardner et al., 2018) and Huggingface (Wolf et al., 2019).

When creating synthetic datasets, it is important to ensure that systematic biases, or annotation artifacts (Gururangan et al., 2018), are not introduced into the resulting probes and that the target datasets are sufficiently challenging (or good, in the sense of Hewitt and Liang [2019]). To test for this, we use several of the MCQA baseline models first introduced in Mihaylov et al. (2018), which take inspiration from the LSTM-based models used in Conneau et al. (2017) for NLI and various partial-input baselines based on these models.

A slight variant of this model, the Choice-to-choice model, tries to single out a given answer choice relative to other choices by scoring all choice pairs $αi,i′(j)=Att(rci(j),rci′(j))∈R$ using a learned attention mechanism Att and finding the choice with the minimal similarity to other options (for full details, see their original paper). In using these partial-input baselines, which we train directly on each target probe, we can check whether systematic biases related to answer choices were introduced into the data creation process.

A Question-to-choice model, in contrast, uses the contextual representations for each question and individual choice and an attention model Att model to get a score $αq,i(j)=Att(rq(j),rci(j))∈R$ as above. Here we also experiment with using ESIM (Chen et al., 2017) to generate the contextual representations for q,c_i (which includes token-wise attention), as well as a VecSimilarity model that measures the average (cosine) vector similarity between question and answer tokens: $αq,i(j)=Sim(embed(q(j)),embed(ci(j)))$⁠. These sets of baselines, which have been shown to be weak on other benchmark MCQA tasks, are primarily used not as competitive models but to check for artifacts between questions and answers that are not captured in the partial-input baselines. This helps ensure that the overall MCQA probing tasks are sufficiently difficult.

Using the various models introduced above, we train these models on benchmark tasks in the science domain and look at model performance on our probes with and without additional training on samples of probe data, building on the idea of inoculation from Liu et al. (2019a). Model inoculation is the idea of continuing to train models on new challenge tasks (in our cases, separately for each probe) using only a small amount of examples. Unlike in ordinary fine-tuning, the goal is not to learn an entirely re-purposed model, but to improve on (or vaccinate against) particular phenomena (e.g., our synthetic probes) that potentially deviate from a model’s original training distribution.

Using this methodology, we can see how much exposure to new data it takes for a given model to master a new task, and whether there are phenomena that stress particular models (e.g., lead to catastrophic forgetting of the original task). Given the restrictions on the number of fine-tuning examples, our assumption is that when models are able to maintain good performance on their original task during inoculation, the quickness with which they are able to learn the inoculated task provides evidence of prior competence, which is precisely what we aim to probe. To measure past performance, we define a model’s inoculation cost as the difference in the performance of this model on its original task before and after inoculation, which serves as a control on the target QA model.

We pre-train on an aggregated training set of all benchmark science exams in Table 6.⁸

In line with our goal of obtaining insights into the strongest QA models, we first pre-trained our RoBERTa-large model on the RACE dataset (Lai et al., 2017), a recipe used by several leading models on science benchmarks. and created an aggregate development set of ∼4k science questions for evaluating overall science performance and inoculation cost. To handle a varying number of answer choices in these sets, we made all sets 5-way by adding empty answers as needed. We also experimented with a slight variant of inoculation, called add-some inoculation, which involves balancing the inoculation training sets with naturalistic science questions. We reserve the MCQL dataset in Table 6 for this purpose, and experiment with balancing each probe example with one science example (x1 matching) and adding twice as many science questions (x2 matching, up to 3k) for each new example.

We use instance-level accuracy, the standard overall accuracy of correct answer prediction (as in Table 7). In addition, we also propose to measure a model’s cluster-level (or strict cluster) accuracy, which requires correctly answering all questions in a semantic cluster (cf. Section 3.1.1).

Our cluster-based analysis is motivated by the idea that if a model truly knows the meaning of a given concept then it should be able to answer arbitrary questions about this concept without sensitivity to varied distractors. Although our strict cluster metric is simplistic, it takes inspiration from work on visual QA (Shah et al., 2019), and allows us to evaluate a model’s consistency and robustness across our different probes, and to get insight into whether errors are concentrated on a small set of concepts or widespread across different clusters.

The ability of a model to answer several questions about a single concept can be thought of as a type of certificate (i.e., further justification and demonstration) of general understanding of that concept in the sense of Ranta (2017).

We begin with an assessment to ensure that our probes are sufficiently difficult to provide meaningful insights into strong models (Section 5.1), then assess the strength of pre-trained QA models (Section 5.2) and whether they can be effectively inoculated (Section 5.3), and finally present a cluster-based consistency analysis (Section 5.4).

Partial-input baseline models, Choice-Only and Choice-to-Choice, generally performed poorly on our probes (cf. Table 7, group 1), indicating limited biases in distractor generation. Initial versions of DictionaryQA had unforeseen biases partly related to distractors sampled from entries without example sentences (cf. Section 3.2), which resulted in high (56%) Choice-Only-GloVe scores before such distractors were filtered out.

One exception is our hypernymy probe where, despite several attempts at filtering data and de-duplicating splits (with respect to correct answer and distractor types), the Choice-to-Choice-BERT/RoBERTa models achieve over 60% accuracy. The nature of the biases here remains unclear, highlighting the importance of having rigorous baselines as unintended biases in expert knowledge can carry over to resulting datasets. We also note the large gap between the BERT/ RoBERTa versus GloVe choice-only models, emphasizing the need for using the best available models even in partial-input baselines.

A more conventional set of Task-Specific QA models (i.e., the LSTM-based Question-to-Choice models trained directly on the probes) is not particularly strong on any of the datasets (cf. Table 7, group 2), suggesting that our probes are indeed sufficiently challenging and largely immune from overt artifacts. The poor performance of the VecSimilarity (which uses pre-trained Word2Vec embeddings without additional training) provides additional evidence of the insufficiency of elementary lexical matching strategies.

Non-transformer science models, such as ESIM with GloVe or ELMo, struggle with all probes (cf. Table 7, group 3), often scoring near random chance. In sharp contrast, the transformer models have mixed results, the most striking being RoBERTa QA models on the definitions, synonymy and hypernymy test probes (achieving 77%, 64%, and 71% respectively), which substantially outperform even task-specific LSTM models trained directly on the probes. Throughout all of these results, however, model performance is significantly behind human performance.

At first glance, these zero-shot results suggest RoBERTa’s high competence on these phenomena. A closer scrutiny enabled by our controlled probes, however, provides a more subtle picture. Each heat map in Figure 3 breaks down the performance of an ESIM or RoBERTa QA model based on the difficulty of the probe dataset (rows) and the nature of the distractors (columns).

Across all datasets and number of hops in the question (i.e., all rows), zero-shot model performance for RoBERTa (bottom-left heat map) is consistently highest among examples with random distractors (the first column) and lowest when distractors are closest in WordNet space (e.g., sister and ISA, or up/down, distractors at distance k′ = 1). For example, RoBERTa’s zero-shot score drops from 88% to 64% when going from random distractors to up/down distractors at k′ = 1.

Further, model performance also clearly degrades for hypernymy and hyponymy as k, the number of hops in the question, increases (see red dashed boxes). For example, the accuracy on questions involving hyponym reasoning with sister distractors of k′ = 1 (column 2) degrades from 47% to only 15% as k increases from 1 to 4. This general tendency persists despite additional fine-tuning, providing evidence of the limited ability of these models to perform multi-hop inference.

How well probe generation templates align with the science training distribution (which we know little about) can significantly impact zero-shot performance (Petroni et al., 2019). Zero-shot results above thus provide a lower bound on model competence on the probed phenomena. We next consider a probe-specific fine-tuning or inoculation step, allowing models to learn target templates and couple this with knowledge acquired during pre-training and science training.

Accuracy after inoculation on 3K probe instances is shown (with inoculation cost in parenthesis) in group 4 of Table 7, for the model with the highest aggregate score on the original task and new probe. Transformer-based models again outperform non-transformer ones, and better models correlate with lower inoculation costs. For example, on synonymy, ESIM’s inoculation cost is 7%, but only ∼1% for BERT and RoBERTa. This emphasizes the high capacity of transformer QA models to absorb new phenomena at minimal cost, as observed earlier for NLI (Richardson et al., 2020).

Figure 4 shows the corresponding learning curves. Transformer QA models learn most tasks quickly while maintaining constant scores on their original tasks (flat dashed lines, plots 1–4), providing evidence of high competence. For BERT and RoBERTa, add-some inoculation (a) improves scores on the probing tasks (solid black and blue lines, plot 1) and (b) minimizes loss on the original task (dashed blue and black lines, plots 2–4).

ESIM behaves quite the opposite (plots 5–6), generally unable to learn individual probes without degrading on its original task. More science data during inoculation confuses it on both tasks.

As the middle-bottom plot of Figure 3 shows, RoBERTa’s performance improves significantly (e.g., from 59% to 77% on 2-hop hyponymy with random distractors) even after inoculation with a mere 100 examples, providing strong evidence of prior competence. After 3k examples, it performs well on virtually all probes. However, results still notably degrade with the number of hops and distractor complexity, as discussed earlier, and we still find its performance to be between 2% and 10% behind human performance.

Table 8 shows mixed results for cluster-level accuracy across the different WordNetQA probes. Our best model is rather robust on the definitions probe. RoBERTa QA’s cluster accuracy is 75%, meaning it can answer all questions correctly for 75% of the target concepts, and that errors are concentrated on a small minority (25%) of concepts. On synonymy and hypernymy, both BERT and RoBERTa are less strong but appear robust on a majority of concepts. In contrast, our best model on hyponymy has an accuracy of only 36%, indicating that the RoBERTa QA models knows only partially about a vast majority of concepts, leaving substantial room for further improvement.

We emphasize that these results only provide a crude look into model consistency and robustness. Recalling dataset details in Table 4, probes differ in terms of the average size of clusters. For example, hyponymy, in virtue of having many more questions per cluster, might simply be a much more difficult dataset for our cluster-based evaluation. In addition, such a strict evaluation does not take into account potentially erroneous questions within clusters, which is an important issue that we leave for future work.

We presented a new methodology for automatically building challenge datasets from knowledge graphs and taxonomies. We introduced several new silver-standard datasets for systematically probing state-of-the-art open-domain QA models. Although our focus was on probing definitions and ISA reasoning, the methodology is amendable to any target knowledge resource or QA domain. We see synthetic datasets and our general methodology as an inexpensive supplement to recent large-scale investment in naturalistic QA dataset construction (Zellers et al., 2018; Sakaguchi et al., 2020) to help better understand today’s models.

We found transformer-based QA models to have a remarkable ability to reason with complex forms of relational knowledge, both with and without exposure to our new tasks. In the latter case (zero-shot), a newer RoBERTa QA model trained only on benchmark data outperforms several task-specific LSTM-based models trained directly on our probes. When inoculated using small samples (e.g., 100 examples) of probing data, RoBERTa masters many aspects of our probes with virtually no performance loss on its original QA task—which we use as a control on the probing quality.

Because these models seem to already contain considerable amounts of relational knowledge, our simple inoculation strategy, which nudges models to bring out this knowledge explicitly while retaining performance on their original task (hence allowing a fairer probe of its knowledge by giving the model the opportunity to learn the probe format), could serve as a simpler alternative to designing new model architectures explicitly encoding such knowledge (Peters et al., 2019).

Regarding our focus on preserving a model’s performance on its original task, one might expect that re-training on relevant knowledge should improve performance. Following other work in this area (Richardson et al., 2020; Yanaka et al., 2020), we found that maintaining performance after additional fine-tuning on specialized datasets is already a tall order given that models are susceptible to over-specialization; indeed, similar issues have been noticed in recent work on large-scale transfer learning (Raffel et al., 2019). We believe that using inoculation for the sole purpose of improving model performance, which is beyond the scope of this paper, would likely require a more sophisticated inoculation protocol. Devising more complex loss functions extending our inoculation strategy to help balance old and new information could help in this endeavor.

The main appeal of automatically generated probes is the ability to systematically manipulate probe complexity, which in turn enables more controlled experimentation as well as new forms of evaluation. It allowed us to study in detail the effect of different types of distractors and the complexity of required reasoning. This study showed that even the best QA models, despite additional fine-tuning, struggle with harder categories of distractors and with multi-hop inferences. For some probes, our cluster-based analysis revealed that errors are widespread across concept clusters, suggesting that models are not always consistent and robust. These results, taken together with our findings about the vulnerability of synthetic datasets to systematic biases and comparison with human scores, suggest that there is much room for improvement and that the positive results should be taken with a grain of salt. Developing better ways to evaluate semantic clusters and model robustness would be a step in this direction.

We emphasize that using synthetic versus naturalistic QA data comes with important trade-offs. Although we are able to generate large amounts of systematically controlled data at virtually no cost or need for manual annotation, it is much harder to validate the quality of such data at such a scale and such varying levels of complexity. Conversely, with benchmark QA datasets, it is much harder to perform the type of careful manipulations and cluster-based analyses we report here. While we assume that the expert knowledge we use by virtue of being hand-curated by human experts, is generally correct by design, we know that such resources are fallible and error-prone. We propose measuring human performance via small samples of probing data, and leave more scalable methods of removing potential noise and adding human annotation to future work.

One of the overarching goals of our approach to model probing is to uncover whether black box models are able to reason in a consistent and correct manner. Our assumption, similar to Clark et al. (2020), is that the ability of a model to mimic the input-output behavior of data generated using expert knowledge gives some evidence of correctness in virtue of such data being correct by construction (see discussion by Ranta (2017)). We emphasize, however, that there are limits to how much we can learn through this type of behavioral testing, given that models are susceptible to exploiting systematic biases in synthetic data and the general difficulty of disentangling a model’s knowledge acquired during pre-training versus fine-tuning (Talmor et al., 2019a). We therefore see efforts to combine behavioral testing with various other analysis methods (Belinkov and Glass, 2019) that aim to uncover correlations and causal patterns between internal model representations and discrete structures (Chrupała and Alishahi, 2019; Vig et al., 2020; Geiger et al., 2020) as a promising direction for future work. This, in combination with extending our probing strategy to other forms of expert knowledge, could prove to be an effective way to engage others working on linguistics and other areas of AI in state-of-the-art NLP research.

We thank the Action Editor and the three anonymous reviewers for their thoughtful comments and feedback. Thanks also to our colleagues at AI2, in particular Peter Clark, Daniel Khashabi, Tushar Khot, Oyvind Tafjord, and Alon Talmor, for feedback on earlier drafts of this work and assistance with various aspects of modeling. Special thanks to Daniel Khashabi for helping with some of the earlier human evaluation experiments.