Reasoning Over Paragraph Effects in Situations

Introduction

Comprehending a passage of text requires being able to understand the implications of the passage on other text that is read. For example, after reading a background passage about how animal pollinators increase the efficiency of fertilization in flowers, a human can easily deduce that given two types of flowers, one that attracts animal pollinators and one that does not, the former is likely to have a higher efficiency in fertilization (Figure 1). This kind of reasoning however, is still challenging for state-of-the-art reading comprehension models. Recent work in reading comprehension has seen impressive results, with models reaching human performance on well-established datasets Devlin et al. (2019); Wang et al. (2017); Chen et al. (2016), but so far has mostly focused on extracting local predicate-argument structure, without the need to apply what was read to outside context.

We introduce ROPEShttps://allennlp.org/ropes, a reading comprehension challenge that focuses on understanding causes and effects in an expository paragraph, requiring systems to apply this understanding to novel situations. If a new situation describes an occurrence of the cause, then the system should be able to reason over the effects if it has properly understood the background passage.

We constructed ROPES by first collecting background passages from science textbooks and Wikipedia articles that describe causal relationships. We showed these paragraphs to crowd workers and asked them to write situations that involve the relationships found in the background passage, and questions that connect the situation and the background using the causal relationships. The answers are spans from either the situation or the question. The dataset consists of 14,322 questions from various domains, mostly in science and economics.

In analyzing the data, we find (1) that there are a variety of cause / effect relationship types described; (2) that there is a wide range of difficulties in matching the descriptions of these phenomena between the background, situation, and question; and (3) that there are several distinct kinds of reasoning over causes and effects that appear.

To establish baseline performance on this dataset, we use a reading comprehension model based on RoBERTa (Liu et al., 2019), reaching an accuracy of 61.6% $F_{1}$. Most questions are designed to have two sensible answer choices (eg. “more” vs. “less”), so this performance is little better than randomly picking one of the choices. Expert humans achieved an average of 89.0% $F_{1}$ on a random sample.

Related Work

There are many reading comprehension datasets Richardson et al. (2013); Rajpurkar et al. (2016); Kwiatkowski et al. (2019); Dua et al. (2019), the majority of which principally require understanding local predicate-argument structure in a passage of text. The success of recent models suggests that machines are becoming capable of this level of understanding. ROPES challenges reading comprehension models to handle more difficult phenomena: understanding the implications of a passage of text. ROPES is also particularly related to datasets focusing on “multi-hop reasoning” Yang et al. (2018); Khashabi et al. (2018), as by construction answering questions in ROPES requires connecting information from multiple parts of a given passage.

The most closely related datasets to ROPES are ShARC Saeidi et al. (2018), OpenBookQA Mihaylov et al. (2018), and QuaRel Tafjord et al. (2019). ShARC shares the same goal of understanding causes and effects (in terms of specified rules), but frames it as a dialogue where the system has to also generate questions to gain complete information. OpenBookQA, similar to ROPES, requires reading scientific facts, but it is focused on a retrieval problem where a system must find the right fact for a question (and some additional common sense fact), whereas ROPES targets reading a given, complex passage of text, with no retrieval involved. QuaRel is also focused on reasoning about situational effects in a question-answering setting, but the “causes” are all pre-specified, not read from a background passage, so the setting is limited.

Recognizing textual entailment

The application of causes and effects to new situations has a strong connection to notions of entailment—ROPES tries to get systems to understand what is entailed by an expository paragraph. The setup is fundamentally different, however: instead of giving systems pairs of sentences to classify as entailed or not, as in the traditional formulation (Dagan et al., 2006; Bowman et al., 2015, inter alia), we give systems questions whose answers require understanding the entailment.

Data Collection

Background passages: We automatically scraped passages from science textbooksWe used life science and physical science concepts from www.ck12.org, and biology, chemistry, physics, earth science, anatomy and physiology textbooks from openstax.org and Wikipedia that contained causal connectives eg. ”causes,” ”leads to,” and keywords that signal qualitative relations, e.g. ”increases,” ”decreases.”We scraped Wikipedia online in March and April 2019. We then manually filtered out the passages that do not have at least one relation. The passages can be categorized into physical science (49%), life science (45%), economics (5%) and other (1%). In total, we collected over 1,000 background passages.

Crowdsourcing questions We used Amazon Mechanical Turk (AMT) to generate the situations, questions, and answers. The AMT workers were given background passages and asked to write situations that involved the relation(s) in the background passage. The AMT workers then authored questions about the situation that required both the background and the situation to answer. In each human intelligence task (HIT), AMT workers are given 5 background passages to select from and are asked to create a total of 10 questions. To mitigate the potential for easy lexical shortcuts in the dataset, the workers were encouraged via instructions to write questions in minimal pairs, where a very small change in the question results in a different answer. Two examples of these pairs are given in Figure 1: switching “more” to “less” results in the opposite flower being the correct answer to the question.

Dataset Analysis

We qualitatively and quantitatively analyze the phenomena that occur in ROPES. Table 1 shows the key statistics of the dataset. We randomly sample 100 questions and analyze the type of relation in the background, grounding in the situation, and reasoning required to answer the question.

We manually annotate whether the relation in the background passage being asked about is causal (a clear cause and effect in the background), qualitative (e.g., as X increases, Y decreases), or both. Table 2 shows the breakdown of the kinds of relations in the dataset.

Grounding

To successfully apply the relation in the background to a situation, the system needs to be able to ground the relation to parts of the situation. To do this, the model has to either find an explicit mention of the cause/effect from the background and associate it with some property, use a common sense fact, or overcome a large lexical gap to connect them. Table 3 shows examples and breakdown of these three phenomena.

Question reasoning

Table 4 shows the breakdown and examples of the main types of questions by the types of reasoning required to answer them. In an effect comparison, two entities are each associated with an occurrence or absence of the cause described in the background and the question asks to compare the effects on the two entities. Similarly, in a cause comparison, two entities are each associated with an occurrence or absence of the effect described in the background and the question compares the causes of the occurrence or absence. In an effect prediction, the question asks to directly predict the effect on an occurrence of the cause on an entity in the situation. Finally, in cause prediction, the question asks to predict the cause of an occurrence of the effect on an entity in the situation. The majority of the examples are effect or cause comparison questions; these are challenging, as they require the model to ground two occurrences of causes or effects.

Dataset split

In initial experiments, we found splitting the dataset based on the situations resulted in high scores due to annotator bias from prolific workers generating many examples Geva et al. (2019). We follow their proposal and separate training set annotators from test set annotators, and find that models have difficulty generalizing to new workers.

Baseline performance

We use the RoBERTa question answering model proposed by Liu et al. (2019) as our baseline and concatenate the background and situation to form the passage, following their setup for SQuAD. To estimate the presence of annotation artifacts in our dataset (and as a potentially interesting future task where background reading is done up front), we also run the baseline without the background passage. Table 5 presents the results for the baselines, which are significantly lower than human performance. We also experiment with first fine-tuning on RACE Lai et al. (2017) before fine-tuning on ROPES.

Human performance is estimated by expert human annotation on 400 random questions with the same metrics as the baselines. None of the questions share the sample background or situation to ensure that the humans do not have an unfair advantage over the model by using knowledge of how the dataset is constructed, e.g. the fact that pairs of questions like in Table 1 will have opposite answers.

Conclusion

We present ROPES, a new reading comprehension benchmark containing 14,322 questions, which aims to test the ability of systems to apply knowledge from reading text in a new setting. We hope that ROPES will aide efforts in tying language and reasoning together for more comprehensive understanding of text.

Acknowledgements

We thank the anonymous reviewers for their feedback. We also thank Dheeru Dua and Nelson Liu for their assistance with the crowdsourcing setup, and Kaj Bostrom for the human evaluation. We are grateful for discussion with other AllenNLP and Aristo team members and the infrastructure provided by the Beaker team. Computations on beaker.org were supported in part by credits from Google Cloud.

References