Neurips_21_22_Limitation_OpenReview/neurips.1026.json

{
  "File Number": "1026",
  "Title": "Decrypting Cryptic Crosswords: Semantically Complex Wordplay Puzzles as a Target for NLP",
  "Limitation": "While the rule-based version includes common clue types, an inherent limitation of this approach is that it is difficult to enumerate all possibilities. For instance, the Deits solver does not include charade-type clues, nor double definitions. Moreover, the rule-based solver uses WordNet’s [8] word similarity functionality to rank outputs, meaning that, in general, it will fail on clues that have definitional parts consisting of more than one word (e.g. “language model” from our example in the introduction would not be matched).",
  "Reviewer Comment": "Reviewer_1: The authors propose a task around cryptic crosswords (of good size, 142K clues, 5K puzzles from the guardian.com). A clue example is \"But everything’s really trivial, initially, for a language model (4)\" , where the answer is BERT. The \"cryptic\" aspect here is that the first letter of the first four words spells out BERT (and thus the phrase \"initially\"). The authors present a good argument (both intuitively and empirically) that this task is challenging for off-the-shelf neural LMs.\nThe authors first describe some carefully constructed splits to test more challenging model generalization (since it appears the naive split allows lots of memorizing). The word-initial disjoint split seems quite interesting.\nThe authors then show that existing baselines generally do not perform well (even on the naive split). They then propose a curriculum learning approach where they add some related datasets (e.g. anagrams etc.) to help implicitly teach the model certain skills that can help with the downstream task. The resulting model outperforms the baselines but it is clear there is a large amount of headroom.\nThe authors not only acknowledge the concurrent work of Efrat et al. 2021 (which proposes a similar task) but also go above and beyond it with an explicit comparison. I think that this is nice.\nLimitations And Societal Impact:\nI don't have any concerns.\nEthical Concerns:\nI don't have any concerns.\nNeeds Ethics Review: No\nTime Spent Reviewing: 1\n\nReviewer_2: Clarity: The paper is well-written and easy to follow. It explains the nuances of the dataset and the motivation for the experimental setup very clearly.\nData and baseline results: This looks like a really difficult task requiring knowledge and reasoning that may not be solved with just more compute. While problems revolving around puns, wordplay, and easier crosswords have been studied before, the nature of the data and the challenges that come with it are novel. It has generally been hard to create NLP benchmarks that are GPT/BERT-proof especially on account of train-test overlap often resulting from random splits and known biases. The benchmark includes disjoint splits designed to better test generalization, and as expected, the T5 baselines performs worse when compared to the random split.\nCurriculum learning method: Examples in the dataset often contain subproblems which can be learned using additional training from other tasks like American-style crosswords and word descrambling. This seems to result in further gains suggesting that T5 can still benefit from such supervision. While the curriculum learning method isn't particularly novel, it is a solid extension of the baselines and adds value to the paper. Given the nature of the paper, I think this is perfectly fine.\nOverall, the experiments are thorough, and suggest that there is significant room for improvement for future work.\nWeaknesses: Human evaluation: It is still unclear exactly how difficult the task is for humans. Presumably, these puzzles appear periodically in the newspaper sources you mentioned and it is reasonable to assume that readers attempt to solve them. How often do they succeed? I understand that such data is likely not available, but it would be nice to see some sort of human evaluation on a smaller sample of the data both to quantify difficulty and for quality control.\nThe paper argues that cryptic crosswords are an excellent domain for developing computational language systems that \"learn and think like humans\". Yet, it is unclear if these puzzles are solvable by non-experts. The paper even seems to suggest that cryptics pose a challenge to novices. Would someone be able to solve these puzzles without significant training? If not, does it make sense to use a rather niche task to develop systems that learn and think like humans? To clarify, I think there is value to studying niche tasks revolving around complex compositional language. It could still tell us interesting things about models and language. I'm unconvinced that they are effective benchmarks for evaluating systems that learn and think like humans.\n(Minor) The cryptic crossword examples used in the paper all revolve around BERT, and were presumably constructed by the authors themselves. While they don't hinder readability (and are fun!), it would be much nicer to see real examples from the dataset.\nLimitations And Societal Impact:\nYes.\nNeeds Ethics Review: No\nTime Spent Reviewing: 1.5\n\nReviewer_3: Review Criteria\nOriginality: The proposed task is new and interesting to NLP practitioners, primarily discussed only in the contemporaneous work by Efrat, et al. (2021). Cryptic crossword clues contain a mixture of linguistic and meta-linguistic phenomena\nQuality: The paper is technically sound, with convincing baselines (e.g., KNN BoW) and targeted experiments (e.g., Sections 6.3-6.4). The paper includes a thorough description of the task domain and motivates future work on cryptic crosswords.\nClarity: The paper clearly describes cryptic crosswords with illustrative examples but would benefit from additional comparison to American-style crosswords, which many readers may be more familiar with. The technical approach is clearly described and replicable.\nSignificance: In my opinion, the proposed task is likely more significant than the proposed curriculum learning approach. Cryptic crossword clues are quite different from standard question-answering benchmarks and may require completely new architectural approaches (e.g., due to segmentation and meta-linguistic reasoning, cf. Section 6.3) from those used in traditional NLP tasks. This is perhaps different from NLU benchmarks like GLUE and SuperGLUE, which track progress in NLP without necessarily requiring specialized model architectures. Additional discussion along these lines would be appreciated.\nNotes on Individual Sections\nIntroduction:\nMaybe worth clarifying that this paper focuses only on questions, rather than full puzzles\nTable 1 is hilarious; I love it. I would normally recommend including real clues rather than synthetic ones, but I think these clues sufficiently convey the conventions of cryptic crosswords. It might also be worth discussing that even though cryptic crossword clues often look impossibly difficult to people who have never seen them before, there's a relatively steep learning curve, such that knowledge of just a few conventions and their associated indicators can make most clues relatively tractable. [This is in contrast to American-style crossword clues, which can mostly be understood without much domain knowledge.]\nRelated Work:\nProbably worth discussing relationship to open-domain question answering\nWhen comparing to American-style crosswords, it might be worth mentioning that solving cryptic crosswords requires additional focus on language and less on search (if you think this is true). Specifically, because cryptic crosswords typically have more black squares than e.g., a standard NYT puzzle, more clues are going to have unique answers. They may also rely less on trivia/general knowledge, etc.\nDataset:\nCurrently, the paper does not include a dataset release but instead provides code for scraping clue-answer pairs from The Guardian website. Releasing the raw data would be better for future-proofing if there aren’t licensing restrictions. I don’t know about The Guardian specifically, but normally full puzzles (including the grid) are copyrighted, but individual clues aren’t. This is why e.g., the OTSYS dataset (see [11]) is permitted.\nCould you report other metrics like MRR or graphs of accuracy@k? While this is perhaps less true for cryptic crosswords vs. American-style crosswords, it’s probably worth noting that top-1 accuracy (or even top-10) is not a great metric for domains with non-unique answers?\nBaselines:\n147: “an inherent limitation of this approach is that it is difficult to enumerate all possibilities” -> I wonder if you could discuss this further? Undoubtedly there is a long tail of linguistic tricks, knowledge, wordplay, etc., but is it really impossible to enumerate all the cryptic crossword conventions? I am not an expert on cryptics by any means, but I have solved a few, and I was under the impression that there is a relatively small number of clue types (maybe 20-30?) with conventionally associated indicators such as “hear” (which indicates homophony).\nI don't have a great intuitive sense of what the Deits solver is doing from this relatively short description. How do the hand-coded grammar and the WordNet components interact? Is one the generator and the other used as a reranker, or is something more nuanced going on?\n167: “leaves considerable room for improvement on the task” -> I would certainly expect that specialized approaches could outperform these numbers, but it’s worth discussing that, especially for top-1 accuracy, the performance ceiling is probably quite a bit lower than 100%.\nSection 4.5: clarify that these numbers are top-1 and top-10 accuracy respectively?\nCurriculum Learning\nIf space permits, maybe include some very brief dataset statistics for ACW?\n\"American crossword clues are usually straightforward synonyms or definitions\" -> can you justify this claim? [I don't think this is true, but there is a domain of \"general knowledge\" crosswords for which it might be, cf. Hill, et al. (2015) \"Learning to Understand Phrases by Embedding the Dictionary\"]\nModel Analysis:\nSection 6.1: how do you specify answer length to the model? Just to make sure, this improves performance even after filtering for length, right? Does this happen only on multiword clues because there is additional information being provided that filtering can't account for? Or e.g., would you expect a similar effect on NYT puzzles where only the full answer length is given?\nSentencePiece tokenization (Sec 6.3) seems like it would be a problem for many of the meta-linguistic conventions in cryptic crosswords; should possibly be mentioned early in the paper, as a sign that the task might not just be a benchmark for current NLP models but instead require specialized architectures\nNits\n13: “a a” -> “a”\nUPDATES AFTER AUTHOR RESPONSE:\nThanks for the very thorough response! The updated information about answer length and tokenization is all very interesting.\nI have increased my score to an 8 following the updated information about the dataset release.\nSorry about the confusion re: evaluation metrics! I had assumed that cryptic crossword clues did not generally have unique answers, akin to American-style crosswords. I think it would be helpful to mention this distinction in the paper as well.\nI agree with other reviewers that Efrat et al. is concurrent and should not affect the decision of this work. Given that there was only one reviewer who supported rejection and their review was primarily based on similarity to Efrat et al., I think the paper should clearly be accepted.\nLimitations And Societal Impact:\nI have two primary concerns about this work:\nDataset Availability and Licensing -- This paper's primary contribution is the proposal of a new task (cryptic crosswords) scraped from The Guardian. However, the authors only release code for scraping the dataset, rather than the data itself. This approach is not future-proof and runs the risk of e.g., link rot and should only be taken if absolutely necessary. Given the discussion of licensing in the checklist, it is not clear to me that the authors are not allowed to release individual clues. Could the authors please provide some additional clarification about this decision to not release the data directly? [In my experience, usually full puzzle grids but not individual clues are copyrighted, which is why the XD and OTSYS datasets are allowed to exist publicly.]\nEvaluation Metrics -- American-style crossword clues do not have unique answers and rely on full puzzle grids to determine uniqueness. Although I am less familiar with UK-style crosswords, I expect the same rule generally holds (although to a lesser extent). Because of this, top-1 and even top-10 accuracy may be inappropriate evaluation metrics for clue-answering, and I recommend including (1) a discussion of whether clues generally have only one acceptable fill and (2) additional evaluation metrics like MRR.\nNeeds Ethics Review: No\nTime Spent Reviewing: 4\n\nReviewer_4: The corpus and task is a compelling contribution. The specifics of the curriculum training are interesting but it is less obvious what their broader contribution will be. A more substantial distinction from the Efrat approach and data is necessary here. This limits the contribution of this work, since it is not the first dataset or the first application of T5 to this task. This indicates that the most substantial contribution is the curricular approach training.\nWhile this is an interesting and challenging task, it is not clear that there is substantial novelty in this work. The corpus, while valuable, is not the first corpus of cryptic clues nor is it clearly preferred over the Efrat corpus. The T5 approach has been explored previously. The major contribution is a training strategy to expose the model to surface manipulations that may help answer certain cryptic clues. A more focused venue might be more appropriate for this work. A more thorough analysis of the types of wordplay that are addressed via the training approach would strengthen the presentation.\nLine 52: typo around the Efrat citation?\nSection 5.1: the descramble and anagrams appear to be most suited to anagram type clues. Are the clues in the data set labeled by wordplay type? If so, does this training help on other clues as well, or only anagrams?\nSection 6.5: is there any reason to not explore the superset of the two corpora together? Is there anything substantially different between the two?\nLimitations And Societal Impact:\nThere is not, to my knowledge, negative societal impact to this work. Though, of course, bias in clueing can be reflected in the model.\nNeeds Ethics Review: No\nTime Spent Reviewing: 2.5",
  "abstractText": "Cryptic crosswords, the dominant crossword variety in the UK, are a promising target for advancing NLP systems that seek to process semantically complex, highly compositional language. Cryptic clues read like fluent natural language but are adversarially composed of two parts: a definition and a wordplay cipher requiring character-level manipulations. Expert humans use creative intelligence to solve cryptics, flexibly combining linguistic, world, and domain knowledge. In this paper, we make two main contributions. First, we present a dataset of cryptic clues as a challenging new benchmark for NLP systems that seek to process compositional language in more creative, human-like ways. After showing that three non-neural approaches and T5, a state-of-the-art neural language model, do not achieve good performance, we make our second main contribution: a novel curriculum approach, in which the model is first fine-tuned on related tasks such as unscrambling words. We also introduce a challenging data split, examine the meta-linguistic capabilities of subword-tokenized models, and investigate model systematicity by perturbing the wordplay part of clues, showing that T5 exhibits behavior partially consistent with human solving strategies. Although our curricular approach considerably improves on the T5 baseline, our best-performing model still fails to generalize to the extent that humans can. Thus, cryptic crosswords remain an unsolved challenge for NLP systems and a potential source of future innovation.",
  "1 Introduction": "Modern computational models have made great progress at handling a variety of natural language tasks that require interpreting rich syntactic and semantic structures [6; 30; 25; 32]. However, in NLP [2; 26; 3] as in other areas of AI [22], machines still lag humans on tasks that require flexible problem solving, rapid learning of unseen tasks, and generalization to new domains. Just as complex games mastered by human experts, such as chess, Go, and video games, have proved a fertile ground for developing more flexible AI [36; 35; 27], we propose that creative language games are a rich area for developing more flexible NLP models. In particular, we argue that linguistic tasks involving metalinguistic reasoning pose an important and significant challenge for state-of-the-art computational language systems.\nOne such domain is cryptic crossword puzzles. Cryptics are the most popular style of crossword in the United Kingdom and appear in major newspapers like The Times and The Guardian. Cryptic clues have two parts: a definition and a wordplay cipher that, when placed adjacent to each other, read like fluent natural language. For example, consider this NLP-centric cryptic crossword clue: “But everything’s really trivial, initially, for a transformer model (4)” with answer “BERT”. “(4)” is the enumeration and specifies the number of letters in the answer. Solvers must identify which\n35th Conference on Neural Information Processing Systems (NeurIPS 2021).\npart of the clue is definition and which is wordplay. The definition should be a semantically valid description of the answer word: “a transformer model” can mean “BERT”. The wordplay part is the remainder of the clue: “But everything’s really trivial, initially”. The word initially in this context is used as an indicator, which tells the solver to take the initial letter of each of the preceding words (but everything’s really trivial) to produce “BERT”. Because both the wordplay and the definition give the same 4-letter word (which is what the enumeration, “(4)”, calls for), we can be confident that “BERT” is the correct answer.\nThe clue just discussed is an example of the initialism clue type, which is one of roughly 15 major clue types. Other types can require anagramming words, finding words hidden in longer phrases, performing synonym substitutions, substituting a word for a soundalike (e.g., “hiring” for “high ring”), or composing a number of these manipulations. See Table 1 for examples of several other clue types with answer “BERT” and Appendix A for examples of actual clues from the dataset. As with American-style clues, cryptics require world knowledge and linguistic flexibility to match the definition, but also considerable attention to meta-linguistic concepts to solve the wordplay. In this paper we study the language task of solving individual cryptic clues, rather than full puzzles since, unlike American-style crosswords, cryptics generally have unique answers and so are less reliant on grid constraints in order to achieve a complete solution.\nWhile cryptics pose a challenge to novice solvers unfamiliar with their structure, experts flexibly combine world, domain-specific, and linguistic knowledge to solve novel clues. Experts know the rules that govern cryptic clues, but they also reason about them flexibly and apply them to solve novel clues. In the psychology literature, it has been claimed that cryptics depend on domain-general intelligence and are an ideal domain for studying the “aha moment” in humans [11; 10] because experts can solve them with high accuracy (but rarely at first glance), they can be easily generated, and they require drawing on a diverse range of cognitive abilities. Therefore, we believe that cryptic crosswords are an excellent domain for developing computational language systems that “learn and think like humans” [22], posing an interesting and important challenge for modern machine learning.\nOur main contributions are, first, a cleaned dataset of cryptic crosswords clues from theguardian.com, consisting of 142,380 clues from 5,518 puzzles over 21 years;1 and second, a novel curriculum learning approach, in which the model is first fine-tuned on related, synthetic tasks (e.g., an augmented word descrambling task) before tackling actual cryptic clues. This method meaningfully improves on a standard T5 seq-to-seq approach and on the best model of Efrat et al. [7]—concurrent work that presents a similar dataset and similar neural baseline using T5.\nIn this paper, we aim not only to present a dataset and propose a novel solution but also to characterize the problem and motivate its importance. To that end, we elucidate the task with three non-neural baselines and fine-tune T5 [31], a Transformer-based [38] encoder-decoder, as a neural baseline. Since the character-level wordplay inherent to cryptics might be challenging to language models with subword tokenization (T5 uses SentencePiece [21]), we study whether T5 has or can acquire metalinguistic knowledge. In Section 6.1 we examine whether T5 learns meta-features of the task related to answer length. In Section 6.3 we use a descrambling task to assess whether T5 understands the character composition of words and whether the model can make use of linguistic and meta-linguistic information simultaneously. Our results show, perhaps surprisingly, that the subword-tokenized T5 model is quite robust to character-level challenges. Moreover, the descrambling task may serve as a useful benchmark task in guiding the development of new approaches on the cryptics task.\nGiven the compositional nature of cryptic clues, we investigate the extent to which the model generalizes under increasingly difficult data splits. In Section 6.2 we introduce a new form of disjoint data split to address T5’s robustness to inflection: a word-initial disjoint split that segments clue– answer pairs based on the first two letters of the answer. In Section 6.4 we examine the systematicity of the model’s answer generation by perturbing the wordplay portion of anagram clues, showing that its behavior is partially consistent with human solving strategies.\nAlthough our novel curricular approach considerably improves performance on this task, fully solving cryptics remains a challenge for modern machine learning, with expert humans still outperforming machines. Therefore, we hope that our dataset will serve as a challenging benchmark for future work.",
  "2 Related work": "While there is an existing literature on puns and wordplay in NLP [17; 20; 24] as well as on solving American-style crosswords [13; 23; 34], there has been relatively little work using NLP methods on cryptics, which require processing highly compositional language and disambiguating surface-level from hidden meaning. Hart and Davis [16] laid out a computational framework for solving the problem, and Hardcastle [14, 15] proposed some rule-based solutions. The app Crossword Genius from Unlikely AI solves cryptic clues and gives human-understandable explanations. Because its method is proprietary and not available for open testing, we do not report it as a baseline but note that it is competitive. Deits [5] offers a rule-based solution, which can also output explanations, and which we test on our dataset. Most recently, in work concurrent to ours, Efrat et al. [7] release a similar dataset of cryptic crossword clues and present a neural baseline using T5-large. In Section 6.5, we discuss how our new data split and curricular approach improve on their work.\nOur curricular approach fits into the space of recent work on pre-finetuning [37; 1] and curricular approaches for compositional tasks [12; 40]. Our approach is loosely related to the pre-finetuning of Aghajanyan et al. [1] but differs in that our curriculum is composed of fewer tasks that are all closely related to the primary task and synthetically generated. Our approach resembles Geva et al. [12] in that we attempt to endow language models with a specific kind of reasoning by training in a multi-task setup over synthetic data. In the vein of Wu et al. [40], which trains on synthetic datasets to encode inductive bias, our approach can be understood as encoding wordplay functional biases.\nWhether large pretrained Transformer models generally pick up on meta-linguistic features like word length and the character composition of words is an open question. Brown et al. [4] explore a set of restricted word unscrambling tasks in the few-shot (versus fine-tuned) setting for GPT-3. Probing results from Itzhak and Levy [18] suggest that information about the character composition of words is present in the embeddings of pretrained, subword-tokenized models. This seems to\n1We release the dataset along with all code to reproduce the results in this paper at https://github.com/ jsrozner/decrypt.\nconfirm our result that subword-tokenized models do have more knowledge of character composition and meta-properties of words than we might have expected.",
  "3 Dataset and task": "We present a cleaned dataset of 142,380 cryptic crossword clues from 5,518 puzzles published in The Guardian from July 1999 to October 2020. We also introduce a challenging “word-initial” disjoint split after observing that T5 is robust to inflection. Overall, the dataset has 55,783 unique answers, giving a mean frequency of 2.55 clues per unique answer. Answers in the dataset consist of one to six words, with most (97.4%) having one (83.3%) or two (14.1%) words. Full details of dataset preprocessing are given in Appendix B, and in addition to releasing the full dataset, we include code to fully replicate our data download, pre-processing pipeline, and split generation in the repository.",
  "3.1 Task": "We frame the problem as a standard seq-to-seq task from inputs (clues with length enumeration) to outputs (answers). For example, one input could be But everything’s really trivial, initially, for a transformer model (4), with output BERT. This is consistent with how clues are presented to human solvers. Full details of the input–output specification for each model are provided in Appendix C, along with other experimental details.",
  "3.2 Splits": "As motivated in the introduction (and in Section 6.2), we consider three splits. The naive (random) split is a standard 60/20/20 random split into train, dev, and test. The disjoint split ensures that all clues with the same answer appear in only one of train, dev, or test. For example, if any clue for which the answer is “BERT” is in the train set, then all clues for which the answer is “BERT” are in the train set. The disjoint split is used to test composition and generalization. It prevents an approach that relies only on lexical similarity across clues (like KNN) from succeeding on the task. Finally, the word-initial disjoint split is designed to address T5’s robustness to inflection. For this split, we enforce that all clues whose answers start with the same two letters will appear in the same set. For example, all clues that have answers starting with ‘ab’ like “abacus,” “abdicate,” “abdicates,” will be grouped, ensuring that inflections or spelling variations of the same base word occur in a single split.",
  "3.3 Metrics": "Our primary metric is whether the top-ranked output is correct, after filtering to outputs of the correct length. We filter because each clue in a puzzle has a hard length constraint, i.e., a human solver cannot pencil in a solution that is of the wrong length. Additionally, we report how often the correct answer is contained in the top 10 outputs after length filtering. This is a useful metric since, when clues are presented in the context of a full puzzle, solvers use information from interlocking answers to narrow down a set of candidates. For instance, the best-performing crossword solver for American-style (non-cryptic) crosswords relies heavily on satisfying grid constraints [13]. Comparing post-length-filter results from Section 4.5 with pre-filter results from Section 6.1, the length filter is seen to increase the top-10 metric by roughly 6% for T5 (with length enumeration given).",
  "4 Baseline models and results": "To characterize how simpler approaches perform on this task, we test three non-neural baselines: a WordNet-based heuristic model, a k-nearest-neighbor bag of words model (KNN BoW), and a rule-based model designed for the task [5]. For a neural baseline, we fine-tune T5-base [31]. For models that can interpret the answer-length enumeration as a textual token (KNN and T5), we append it to the input string (e.g., “(4)”). For these two models, we report results both with and without appending the enumeration. Implementation details are discussed in Appendix C.",
  "4.1 WordNet": "Our first baseline is a simple heuristic based on clue structure. It takes advantage of the fact that the definition part of a cryptic clue always appears at the beginning or end of the clue. For instance, in the double definition clue for “BERT” in Table 1, “Model Sesame Street character,” the word “model” appears at the beginning and is a definition (in this case, a hypernym) for the answer “BERT”. We use WordNet [8], a large database of English word meanings, to build a reverse dictionary via the synonym, hyponym, and hypernym graphs. We take as candidate solutions the set of reverse dictionary outputs for the first and last words of a clue. For example, if “dog” appears at the start of a clue, candidates would include “animal”, “canine”, “labrador”, etc. Ranking outputs by characterlevel overlap with the rest of the clue slightly improves performance, since answers sometimes are rearrangements of the characters in the wordplay portion of the clue.",
  "4.2 KNN BoW": "To assess whether clues that are close in lexical space have similar answers, we use a KNN model on top of a bag-of-words featurizer.",
  "4.3 Rule-based": "Finally, to evaluate how well a rule-based approach, with a hand-coded grammar, can perform on the task, we use the Deits [5] solver.2 This solver handles anagrams, initialisms, substrings, insertions, and reversals in a rule-based way. While the rule-based version includes common clue types, an inherent limitation of this approach is that it is difficult to enumerate all possibilities. For instance, the Deits solver does not include charade-type clues, nor double definitions. Moreover, the rule-based solver uses WordNet’s [8] word similarity functionality to rank outputs, meaning that, in general, it will fail on clues that have definitional parts consisting of more than one word (e.g. “language model” from our example in the introduction would not be matched).",
  "4.4 T5: vanilla seq2seq": "For our baseline neural seq-to-seq approach, we fine-tune the Transformer-based [38] T5-base model [31], starting from HuggingFace’s [39] pretrained model parameters. T5 is an encoder-decoder language model pretrained on the C4 corpus [31]. Fine-tuning is done via supervised learning (teacher-forcing) over standard seq-to-seq input-output pairs. At test time, we generate outputs using beam search. As described in Section 3.3, we filter the outputs to those of the correct length and evaluate by checking whether the top result is correct and whether the answer occurs in the top 10. See Appendix C for details, including hyperparameter selection.",
  "4.5 Results": "In Table 2, we report metrics for dev and test sets on both the naive (random) split and word-initial disjoint split (discussion on disjointness in Section 6.2). While the WordNet baseline achieves some success (2.6% and 10.7% top-1 and top-10 on the test set), it is inherently limited, since it cannot handle clues with multiword definitions and lacks a good ranking mechanism. KNN does better, achieving 6.1% and 11.3% with length enumeration. The rule-based solver achieves 7.3% and 14.7%, marginally outperforming the KNN baseline. Though our T5 neural baseline outperforms all non-neural baselines, achieving 16.3% and 33.9%, it leaves considerable room for improvement.",
  "5 Curriculum learning": "Solving cryptic clues uses different linguistic and reasoning abilities compared to the natural language tasks on which T5 is pretrained. Thus, during fine-tuning on the cryptics task, the model must learn many sub-parts of the problem simultaneously: how to look up definitions, how to identify wordplay indicators, how to perform wordplay operations, etc. Although these elements of the task\n2Deits [5] has a more recently implemented solver in Julia that was used by Efrat et al. [7]. We use the Python version, which may have small differences from the Julia version and is reportedly much slower (see Appendix C for more details).\neach individually benefit from T5’s natural language pretraining, our standard T5 baseline suggests that learning to compose them all at once is challenging. We show that a curriculum of synthetic datasets can address this, substantially improving performance on the primary cryptics task. We test a number of different curricular tasks and task sets and discuss which are most effective. This curricular approach may be useful in other settings where focused linguistic capabilities are required.",
  "5.1 Curricular datasets": "The logic of our curricular approach is to provide some guidance on the sub-parts of the problem space before building up to fully compositional cryptic clues. For curricular training, we create four datasets designed to improve performance and elucidate what makes a good curricular task for this problem. We process a public American crossword clue dataset [29], henceforth “ACW-data” for “American Crossword”, and generate three datasets: ACW, ACW-descramble, ACW-descramble-word. After preprocessing, ACW-data has 2.5m clue–answer pairs with 250k unique answers, giving a mean frequency of roughly ten clues per unique answer. Unlike cryptic clues, American crossword clues often involve relatively straightforward synonym or definition substitions, so the ACW dataset can be used to train definition lookup.\nWe also produce a separate anagram dataset from a publicly available English dictionary. Details to produce all datasets are included in Appendix D.1. In all example clues that follow, the target output is “PETAL” and we use labels (prepending a word and colon) to help the model distinguish tasks.\n1. ACW: ACW-data dataset in input–output form (i.e., a seq-to-seq version of American-style crossword clues) with no augmentation. For example, phrase: flower part (5).\n2. ACW-descramble: For each clue–answer pair in ACW-data, we create an input that models a cryptic clue by scrambling the answer word and prepending or appending it to the clue portion. For example, we scramble “petal” and randomly place it at the beginning (descramble: etalp flower part (5)) or end (descramble: flower part etalp (5)) of a clue.\n3. ACW-descramble-word: A seq-to-seq task that is just the descrambling of answer words. When compared to ACW-descramble, this elucidates the importance of curricular and primary task similarity, in particular whether teaching a bare descrambling task is helpful to the model. Example: descramble word: etalp (5).\n4. Anagrams: Using a dictionary as starting point, we synthesize an anagram dataset: we pair a word (to be anagrammed) with an anagram indicator (e.g., “mixed up”, “drunk”, “confusingly”) and ask the model to produce a valid anagram (i.e., a scrambled version of the word that is itself also a valid word). For example, anagram: confusingly plate (5) (rearrange the letters of ‘plate’ to get “PETAL”). The anagram dataset simulates the anagram type of wordplay in cryptic clues, with definition part omitted.\nTable 3: Curricular results (left) and sample metrics for meta-linguistic feature analysis (right)\n(a) Curricular approaches on the naive (random) split. Metric is correctness of top-output (5 beams with length filter).\nPercent correct Curricular dataset full dev\nset anagram subset\nBaseline (no curricular) 15.7 13.7\nACW 18.3 14.4 ACW-descramble 13.1 21.4\nACW + ACW-descramble 20.2 24.0 ACW + ACW-descrambleword 17.8 18.3\nACW + anagram 17.1 19.1 ACW + ACW-descramble + anagram 20.1 27.1\n(b) Sample metrics calculated over top 10 outputs without length filter, using naive split.\nModel % samplecontains answer\n(top-10, no filter)\n% outputs with correct\nlength\n% outputs correct\nword count\ndev test dev test dev test\nKNN – (no lengths) 6.5 6.6 13.4 13.3 70.7 70.7 – (lengths) 10.6 10.7 85.4 85.3 89.7 89.6\nT5-base – (no lengths) 19.0 18.8 16.0 16.2 74.2 74.1 – (lengths) 27.5 28.1 48.3 48.5 97.9 97.9",
  "5.2 Methods": "Our curricular approach is as follows: using the same methods as in Section 4.4, first, we fine-tune T5 on one or more supplementary tasks. Second, we continue fine-tuning on the primary cryptics task, periodically showing batches from the supplementary task(s), to decrease the likelihood of catastrophic forgetting [9]. Full details of our curricular training setup are in Appendix D.2.",
  "5.3 Results": "In Table 3a, we report results for our curricular approach. Overall, we find that curricular training using ACW + ACW-descramble is best and report in Table 2 a primary task improvement from 16.3% to 21.8% on the random split and from 1.1% to 6.5% on the word-initial disjoint split.\nWe begin by testing a curriculum with only ACW, which corresponds roughly to teaching the definitional lookup. Observing that ACW improves performance over the baseline, we test ACWdescramble, which is ACW augmented with a scrambled word component (an implicit wordplay). Surprisingly ACW-descramble leads to a decline in performance relative to both ACW and to Baseline. On the other hand, combining ACW + ACW-descramble leads to our best performing approach, demonstrating that a particular curricular combination can improve over two separate tasks. We also compare ACW + ACW-descramble to ACW + ACW-descramble-word. The drop in performance suggests that inputs that are more similar to the final task are more useful than, e.g., a bare descrambling task. To isolate the effect of task choice, all curricular approaches use the same data distribution, train for the same number of curricular fine-tuning steps, and are reviewed during primary training at the same frequency.\nFinally, in order to explore whether the curricular training improves performance across the board or just on the clue types directly involved (e.g., anagrams), we report performance on an algorithmicallylabeled anagram subset (“anagram subset” column in Table 3a). Adding the Anagrams subtask to the curriculum improves anagram performance but interestingly does not improve performance compared to the top curriculum, ACW + ACW-descramble.3 We see a similar gain in anagram performance (but drop in overall performance) when training with only ACW-descramble. This suggests that pretraining for a particular clue type can improve performance on that type, but perhaps at the expense of performance on other clue types.\nBeyond providing guidance to T5 on the problem sub-parts, this approach also partially addresses the disjointness of train and test sets. For the word-initial disjoint split, by periodically refreshing the curricular task, we remind T5 to produce outputs that are not in the training set of the cryptic split.\n3The distribution and size of the Anagrams dataset is different from the ACW datasets, so we separate curricula involving the Anagrams dataset in Table 3a.",
  "6.1 Learning meta-linguistic properties: output length and number of words": "This task requires not just linguistic but also meta-linguistic reasoning. We investigate how model behavior changes when the enumeration (the specification of the length of the answer) is appended to the end of the input string as a number in parentheses. Whether large pretrained transformer models generally pick up on meta-linguistic features like word length is an open question.\nTo study whether the models learn length information, we report how often the top-10 candidate outputs for each clue are the correct length and have the correct number of words before applying any length filter. For both the KNN and the T5 models, we find that including the length enumeration improves overall performance, as can be seen in Table 2. (We omit the WordNet and rule-based approaches from this discussion since they have no capacity to learn the meaning of the length enumeration.)\nIn columns 3 and 4 of Table 3b we see that both KNN and T5 pick up on length information, generating more outputs of the correct length when the enumeration is provided. T5 is particularly proficient at producing the correct number of words in outputs. (Recall that multiword answers are indicated with an enumeration that has two numbers separated by a comma, as in “(4, 6)”, indicating an answer like “ALAN TURING”.) Given that T5 produces 97.9% of outputs with the correct number of words, it seems plausible that the model is learning a correspondence between the enumeration and the presence of a space or spaces in its output.",
  "6.2 Disjointness": "Based on the performance of the KNN model on the naive data split, we see that some clues can be solved by picking up on lexical similarity to other clues. Thus, we investigate whether T5 is also picking up on similarity to previously seen clues or if it is learning something about the compositional and functional structure of the cryptic clues.\nTo assess how much a Transformer-based model like T5 relies on having seen similar clues for the same word, we segment performance on the random split by whether the answer was in the train set. In Table 4a, we see that performance drops from 16% on the full dev set to only 3.0% on the clue subset not seen during training, confirming our intuition that lexical similarity between clues with the same answer plays a role in model success.\nTo formalize this result, we create and train on the two disjoint datasets described in Section 3: the basic disjoint and the word-initial disjoint splits. The T5 model achieves 3.3% accuracy on the basic disjoint split (dev) and only 1.1% accuracy on the word-initial disjoint split. The drop is likely partially attributable to robustness to inflection, since inflected forms often start with the same two letters.",
  "6.3 Wordplay: minimal task": "We investigate the extent to which T5, which uses SentencePiece tokenization [21], can perform wordplay-esque tasks like descrambling. We run an experiment as follows: we start with the ACW dataset from Section 5, further restrict to outputs with targets in a dictionary (i.e., no multiword answers), and downsample to 180k clues (10%). We create two descrambling tasks. The first is a direct descramble task, where the input is a scrambled version of the target word (e.g., etalp for target petal). The second task is a descramble with phrase tasks, in which we append the clue of the clue-answer pair after the scrambled answer letters (e.g., input is etalp | flower part for target petal). The second task is designed to mimic the cryptic setup, where we have a wordplay (in this case, the implicit descramble function) whose output is conditioned on a (possibly phrasal) synonym. See Appendix E.1 for more details.\nWe present results in Table 4b. We observe that the model reasonably learns the task on a random split (63.8%) but fails on a word-initial disjoint split (3.7%). Notably, including a phrasal definition alongside the scrambled letters in the input improves outcomes, suggesting that the model is simultaneously incorporating both meta-linguistic character-level and overall word-embedding information. This task can serve as a benchmark for models to solve the cryptic task, since it roughly upper-bounds how well a model can solve wordplays.\nTable 4: Disjointness results (left) and descrambling results (right)\n(a) T5-base performance (% for top-10 outputs after length filter) on naive, subset of naive not seen in train, disjoint, and word-initial disjoint splits.\nDataset Top Top 10\ncorrect contains dev test dev test\nNaive (random) split – Entire split 16.0 16.3 33.1 33.9 – Subset not in train 3.0 2.8 9.5 9.7\nDisjoint splits: – Naive disjoint 3.3 3.2 12.6 12.9 – Word-initial disjoint 1.1 1.1 5.6 5.8\n(b) Descrambling task, with and without phrasal definition. Metric is % for top ten outputs without length filter.\nSplit and task Top Top 10 correct contains\nRandom split – Descramble 63.8 91.6 – Descramble w/ phrase 79.4 91.2\nWord-initial disjoint split – Descramble 3.7 12.5 – Descramble w/ phrase 12.4 24.4\nGiven that we observe a considerable drop from the random to disjoint data splits, we test whether T5 can learn the identity function under a disjoint split. We find that the model achieves 100% accuracy on a direct copy task in which the model is trained to simply output the string that is given as input. This suggests that T5’s meager performance for descrambling on the disjoint split is not due to an inability to generate an output that has not been seen in fine-tuning.",
  "6.4 Assessing systematicity": "We investigate the extent to which the model’s behavior is consistent with the compositional and systematic nature of human solving strategies. Consider our anagram clue from Table 1: “Confused, Bret makes a language model (4),” for which the answer is “BERT”. Using a publicly available list of 18,000 first names, we algorithmically identify clues in the dev set that require the solver to compute an anagram of a first name, and we use those clues to compose two specialized test sets. In the first set, we scramble the original letters (e.g., Bret becomes Treb). If the model is behaving optimally, performance on this set should not suffer: what is important about Bret in this clue is not its semantics but its characters. In the second set we substitute another name of the same length (e.g., John for Bret). We pick first names because swapping a name does not change the grammaticality or linguistic validity of clues. Here, for a human solver, we would expect a major hit in performance since the correct answer is a valid anagram of Bret but not of John, and so the clue is no longer a valid cryptic clue. It does, however, still have a valid definition part (“a language model”), and so, if the model is picking up only on the definition part and ignoring the wordplay part, it might still do well. See Appendix E.2 for more task details.\nOn this set of clues, prior to any modification, we find a baseline accuracy of 33.3%. When scrambling the name, accuracy drops moderately, to 25.2%. When substituting a name with different letters, accuracy falls to 7.0%. This difference in performance on the substitution and scrambling tasks suggests that the model is, to some extent, correctly picking up on the need to use the letters of the name as the wordplay substrate. This is confirmed by the average character overlap between the altered word and the generated candidate answers. We observe 51.2% multiset overlap for the baseline (name unchanged from original clue), 51.0% for substitution of names with the same letters, and 31.4% when substituting names with different letters. For our example clue, this means that we would expect high character overlap between Bret and the candidate answers in the baseline set, but high overlap between John and the candidate answers in the substitution test set. These results suggest that, like human solvers, the model is sensitive to character manipulations at the location of the wordplay substrate.",
  "6.5 Comparison to Efrat et al.": "In contemporaneous work, Efrat et al. [7] present a dataset of cryptics from two other major newspapers and fine-tune T5-large for the task. While Efrat et al. [7] conclude that train/test disjointness is important, they do not fully consider T5’s robustness to plural and other inflections. The word-initial disjoint split that we present addresses this. In particular, their ‘naive’ split is the same as our naive split, and their ‘official’ split is the same as our (naive) disjoint split. To demonstrate that our split is\nrelevant to the Efrat work, we replicate their results (we train T5-large and report the same metric, correctness of top output with b=5 beams), show a considerable decline in performance under the word-initial disjoint split (10.9% to 4.9%), and finally demonstrate that our curricular approach substantially improves results on the ‘naive-disjoint’ (10.9% to 19.9%) and word-initial disjoint splits (4.9% to 12.8%). Performance on the ‘naive’ split does not change considerably with our curricular approach. Results are in Table 5, and further training details are given in Appendix E.3.",
  "7 Conclusion": "In this paper we introduce a dataset of cryptic crossword clues that can be used as a challenging new benchmark task and develop a novel curricular approach that considerably improves performance on the benchmark. We further characterize the problem with three non-neural baselines and provide methods for investigating model behavior, including a simple descrambling task and an experiment that explores what T5 learns about compositional task structure. Lastly we introduce a challenging word-initial datasplit to evaluate a model’s ability to achieve compositional generalization. These contributions demonstrate why this task is worth studying and how it may be relevant to related problems. For example, our curricular approach may be useful in other settings where focused linguistic capabilities are required.\nPretrained contextual embedding models like T5, which draw on a rich base of lexical and linguistic knowledge from pretraining, are a promising candidate for the type of flexibility needed to solve this sort of puzzle. However, T5 does not initially succeed at the task, and although our curricular approach considerably improves task performance, cryptics remain an unsolved problem. Although one might initially think that a character-level tokenization scheme would be necessary for this task, Sections 6.1 and 6.3 suggest that T5 can unscramble words (under appropriate generalization splits) and does learn a correspondence between a word’s tokens and the word’s total length.\nGiven the success of our curricular approach, future work might combine new synthetic datasets under a learned curriculum schedule. In any case, an approach that fully solves this problem will need to more flexibly learn different kinds of wordplay functions and how to functionally compose them to produce the answer word. In that sense, we believe that the cryptic crossword task serves as a good benchmark for those interested in building NLP systems that can apply linguistic and meta-linguistic knowledge in more creative, flexible, and human-like ways.\nAcknowledgments and Disclosure of Funding\nThis work was supported in part by a Stanford HAI Hoffman–Yee grant.\nWe thank Saul Pwanson for guidance on the release of crossword-related datasets and William Tunstall-Pedoe for discussions on cryptics and for testing his proprietary solver on our dataset. We thank Armando Solar-Lezama, Josh Tenenbaum, Osbert Bastani, and other members of the Neurosym group for helpful feedback. We especially thank Dana Angluin for suggesting (in 2014!) cryptic crosswords as an undergraduate thesis topic for Joshua Rozner.",
  "Reviewer Summary": "Reviewer_1: Interesting task that could be a good challenge for neural models. I generally favor accepting new resource papers especially ones that present challenging tasks.\n\nReviewer_2: The paper focuses on data and methods for studying compositional language through UK-style cryptic crossword puzzles. These puzzles contain a definition and a wordplay cipher which together read like fluent natural language. The output is a word or phrase related to the wordplay cipher for which the definition is a valid description. Expert humans typically use creativity and linguistic and world knowledge to solve this task. The paper makes two main contributions: (1) a new dataset of cryptic clues as a new benchmark for NLP systems that process compositional language, and (2) baselines and an improved curriculum learning approach in which the model is first fine-tuned on related tasks (e.g., unscrambling words). Their experiments suggest that there is significant room for improvement for future work on this new benchmark.\n\nReviewer_3: This paper provides a new question-answering benchmark composed of cryptic crossword clues and describes a curriculum learning-based approach to this task.\n\nReviewer_4: This paper describes a new corpus of cryptic crossword clues as a challenging NLP task. Cryptic crossword clues are composed of a literal (synonym) clue and wordplay. Cryptic wordplay has a conventionalized inventory of ways to clue an answer. This is a compelling NLP challenge ``1) because it is hard and 2) because it requires a system to process both surface, syntactic and semantic content of the examples. This is related to being able to understand statements like \"Al and Bob got married and bought a house together, but not in that order\". The paper includes a curriculum training approach where a T5 model is presented with data to improve cryptic clue solving."
}