Base_JSON/prefixS/json/scil/2020.scil-1.6.json

{
    "paper_id": "2020",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T15:38:46.302710Z"
    },
    "title": "Neural network learning of the Russian genitive of negation: optionality and structure sensitivity",
    "authors": [
        {
            "first": "Natalia",
            "middle": [],
            "last": "Talmina",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Johns Hopkins University",
                "location": {}
            },
            "email": "talmina@jhu.edu"
        },
        {
            "first": "Tal",
            "middle": [],
            "last": "Linzen",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Johns Hopkins University",
                "location": {}
            },
            "email": "tal.linzen@jhu.edu"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "A number of recent studies have investigated the ability of language models (specifically, neural network language models without syntactic supervision) to capture syntactic dependencies. In this paper, we contribute to this line of work and investigate the neural network learning of the Russian genitive of negation. The genitive case can optionally mark direct objects of negated verbs, but it is obligatory in the existential copula construction under negation. We find that the recurrent neural network language model we tested can learn this grammaticality pattern, although it is not clear whether it learns the locality constraint on the genitive objects. Our results further provide evidence that RNN models can distinguish between optionality and obligatoriness.",
    "pdf_parse": {
        "paper_id": "2020",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "A number of recent studies have investigated the ability of language models (specifically, neural network language models without syntactic supervision) to capture syntactic dependencies. In this paper, we contribute to this line of work and investigate the neural network learning of the Russian genitive of negation. The genitive case can optionally mark direct objects of negated verbs, but it is obligatory in the existential copula construction under negation. We find that the recurrent neural network language model we tested can learn this grammaticality pattern, although it is not clear whether it learns the locality constraint on the genitive objects. Our results further provide evidence that RNN models can distinguish between optionality and obligatoriness.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Statistical language models are probability distributions over sequences of words, which they learn from large corpora during training. For any given context, these models assign a probability to all of its possible continuations: for a example, given the context \"he was eating soup with a. . . \", language models can predict that the word \"spoon\" is much more likely to occur next than \"shoe\".",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "A class of language models -Recurrent Neural Network (RNN) models -have been particularly successful on various applied language tasks (Mikolov et al., 2010; Vinyals et al., 2015; Kiperwasser and Goldberg, 2016; Bahdanau et al., 2014) . But what kind of linguistic knowledge do these models capture? Arguably, human language knowledge is comprised of more than word co-occurrence statistics -it encompasses abstract rules and generalizations that concern hierarchical structure. According to the argument from the poverty of the stimulus (Chomsky, 1980) , the kind of structural knowledge that underlies hu-man linguistic performance is impossible to derive purely from the input language learners receive, since many structure-dependent linguistic phenomena are too infrequent in the type of input humans encounter during language acquisition. Therefore, according to the argument, human sensitivity to the structure in language must be innate.",
                "cite_spans": [
                    {
                        "start": 135,
                        "end": 157,
                        "text": "(Mikolov et al., 2010;",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 158,
                        "end": 179,
                        "text": "Vinyals et al., 2015;",
                        "ref_id": "BIBREF23"
                    },
                    {
                        "start": 180,
                        "end": 211,
                        "text": "Kiperwasser and Goldberg, 2016;",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 212,
                        "end": 234,
                        "text": "Bahdanau et al., 2014)",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 538,
                        "end": 553,
                        "text": "(Chomsky, 1980)",
                        "ref_id": "BIBREF5"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Since neural networks do not possess this innate bias -but perform applied natural language tasks with high accuracy -they can provide a rich source of information about the mechanisms underlying hierarchical structure rule learning. A number of questions need to be asked. How much grammar can language models learn just from a corpus? What are the limitations on the generalizations they can make about hierarchical structures? Recently, several studies have addressed these questions by testing RNNs' performance on structure-sensitive grammatical tasks. The results of these studies showed that RNNs can learn subject-verb agreement (Linzen et al., 2016; Gulordava et al., 2018; Ravfogel et al., 2018) , fillergap dependencies (Wilcox et al., 2018) , hierarchical rules of question formation (McCoy et al., 2018) , and the contexts that license negative polarity items (Jumelet and Hupkes, 2018) .",
                "cite_spans": [
                    {
                        "start": 637,
                        "end": 658,
                        "text": "(Linzen et al., 2016;",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 659,
                        "end": 682,
                        "text": "Gulordava et al., 2018;",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 683,
                        "end": 705,
                        "text": "Ravfogel et al., 2018)",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 731,
                        "end": 752,
                        "text": "(Wilcox et al., 2018)",
                        "ref_id": "BIBREF24"
                    },
                    {
                        "start": 796,
                        "end": 816,
                        "text": "(McCoy et al., 2018)",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 873,
                        "end": 899,
                        "text": "(Jumelet and Hupkes, 2018)",
                        "ref_id": "BIBREF10"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In this paper, we contribute to this line of research by extending it to issues in Russian syntax. What makes Russian compelling is that it has rich morphology, which allows us to expand the range of tasks that have been used in previous work to explore RNN learning of structural dependencies. In particular, Russian has casemarking alternations involving the genitive case: along with the accusative case (which is typical cross-linguistically), the genitive can mark direct objects of transitive verbs. However, it is only licensed under negation, and is optional -the accusative case can be used in both affirmative and negative clauses. The genitive also alternates with the nominative case to mark the subjects of existential copula constructions, where it is obligatory under negation. Nominative subjects are only allowed with affirmative sentences. We spell out these properties in more detail in the next section.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "2 Background: Russian genitive-of-negation",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In Russian, direct objects are usually marked by the accusative case, as is common in languages with overt case marking:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "(1) Uchitel Teacher proveril graded domasniye zadaniya homeworksACC \"The teacher graded the homeworks.\"",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "However, non-oblique arguments can receive genitive case in the scope of sentential negationa phenomenon known as the genitive of negation (Bailyn, 1997; Pesetsky, 1982; Paducheva, 2004; Harves, 2002; Timberlake, 1975; Babby, 1980) : If the sentence is affirmative, only the accusative case can be used to mark the direct object: Further, the genitive is only licensed when the negation term is local: in sentences like (5), the relative clause negation cannot license genitive case-marking on the main verb object domasnih zadaniyj. We will refer to this licensing pattern as the LOCALITY CONSTRAINT. \"The teacher, who didn't like the students, graded the homeworks.\"",
                "cite_spans": [
                    {
                        "start": 139,
                        "end": 153,
                        "text": "(Bailyn, 1997;",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 154,
                        "end": 169,
                        "text": "Pesetsky, 1982;",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 170,
                        "end": 186,
                        "text": "Paducheva, 2004;",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 187,
                        "end": 200,
                        "text": "Harves, 2002;",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 201,
                        "end": 218,
                        "text": "Timberlake, 1975;",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 219,
                        "end": 231,
                        "text": "Babby, 1980)",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "The genitive of negation is considered to be optional in sentences like (3) (Kagan 2010 , although see Bailyn 1997 Harves 2002 for discussion), but it is obligatory in the existential copula construction, where the genitive alternates with the nominative case: ",
                "cite_spans": [
                    {
                        "start": 76,
                        "end": 87,
                        "text": "(Kagan 2010",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 88,
                        "end": 114,
                        "text": ", although see Bailyn 1997",
                        "ref_id": null
                    },
                    {
                        "start": 115,
                        "end": 126,
                        "text": "Harves 2002",
                        "ref_id": "BIBREF9"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "(6) (Bailyn,",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Motivated by the observations in the previous section, we explored how well language models can capture the properties of the genitive of negation. We ran a series of experiments to study the behavior of an RNN language model trained by Gulordava et al. (2018) . In Experiment 1, we tested the language model on simple sentences with case-marking alternation on direct objects, finding that the model learned the grammaticality pattern in (3-4). In Experiments 2-4, we tested whether the model was sensitive to the structurally defined scope of negation. We found that the model correctly predicted the genitive-accusative alternation even when there was no overt marking of sentential scope. In Experiment 5, we tested the model on the existential copula construction in which the genitive case is obligatory under negation. Our results suggest that the model could differentiate between the syntactic structures where the genitive case is obligatory from those where it is optional.",
                "cite_spans": [
                    {
                        "start": 237,
                        "end": 260,
                        "text": "Gulordava et al. (2018)",
                        "ref_id": "BIBREF7"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Overview of experiments",
                "sec_num": "3"
            },
            {
                "text": "To explore whether RNN language models can capture the constraints on genitive-marked direct objects, we studied the performance of the model presented in Gulordava et al. (2018) . The model was trained on a 90-million-word corpus extracted from the Russian Wikipedia and had two layers of 650 hidden LSTM units. Additionally, we trained a 3-gram model on the same corpus to provide a baseline for our experiment. The 3-gram model which backs off to smaller n-grams using linear interpolation.",
                "cite_spans": [
                    {
                        "start": 155,
                        "end": 178,
                        "text": "Gulordava et al. (2018)",
                        "ref_id": "BIBREF7"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methodology",
                "sec_num": "4"
            },
            {
                "text": "Following previous work (Linzen et al., 2016; Gulordava et al., 2018; Marvin and Linzen, 2018) , we assessed the model's performance by examining the probabilities it assigned to grammatical sentences from our dataset, compared to ungrammatical ones. We used surprisal (Hale, 2001) :",
                "cite_spans": [
                    {
                        "start": 24,
                        "end": 45,
                        "text": "(Linzen et al., 2016;",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 46,
                        "end": 69,
                        "text": "Gulordava et al., 2018;",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 70,
                        "end": 94,
                        "text": "Marvin and Linzen, 2018)",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 269,
                        "end": 281,
                        "text": "(Hale, 2001)",
                        "ref_id": "BIBREF8"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methodology",
                "sec_num": "4"
            },
            {
                "text": "surprisal(w i ) = log P(w i | w 1 . . . w i 1 )",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methodology",
                "sec_num": "4"
            },
            {
                "text": "The higher the surprisal, the more unexpected a word is under the model's probability distribution. Since the sentences in (3) and (4) are minimally different from each other (the only difference being that the verb in (3) is negated), we can directly compare the surprisal the model assigned to the genitive-marked objects in these sentences. Assuming the probability distribution defined by the model reflects the grammar of the genitive of negation construction, we expected that the genitive-marked object would be assigned higher surprisal in (4), where it is not licensed by negation. Since accusative objects are grammatical independently of polarity, we did not expect the same difference between (1) and (2).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methodology",
                "sec_num": "4"
            },
            {
                "text": "5.1 Experiment 1: Simple sentences",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "5"
            },
            {
                "text": "We constructed a dataset of 64 sentences, each consisting of a subject, a verb, and an object. For each sentence, we included four versions which varied in main verb polarity (positive or negative) and the case marking of the direct object (accusative or genitive), yielding a total of 256 experimental items. Examples (7a-7d) represent all four conditions for one item in our dataset. Only the sentence in (7b) is ungrammatical: both (7a) and (7c) are grammatical because accusative objects are always licensed, and in (7d), the genitive of negation is grammatical because it is within the scope of a negated verb. In (7b), however, the genitive-marked object is not licensed by negation, which makes the whole sentence ungrammatical. Given this pattern, we expected that the model would assign higher surprisal to the word provala 'failure.GEN' in (7b) than in (7d), but there would be no such difference for the word proval 'failure.ACC' in (7a) and (7c).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Materials",
                "sec_num": "5.1.1"
            },
            {
                "text": "LSTM Consistent with our predictions, the genitive-marked direct objects were less surprising when the verb was negated (see Figure 2a) . Figure 3a shows that the difference between the positive and negative conditions is much bigger for genitive-marked objects than for the accusativemarked ones. This suggests the model learned that the negative-polarity constraint only applies to objects marked by the genitive case.",
                "cite_spans": [
                    {
                        "start": 138,
                        "end": 147,
                        "text": "Figure 3a",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 125,
                        "end": 135,
                        "text": "Figure 2a)",
                        "ref_id": "FIGREF6"
                    }
                ],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.1.2"
            },
            {
                "text": "We further tested this by running a linear mixed effects model (Baayen et al., 2008) with the model-assigned surprisal as the dependent variable, and case, polarity, their interaction, and item frequency as predictors. We found a main effect of case (p = 0.004), as well as an interaction between case and polarity (p < 0.0001). Surprisal was significantly affected by polarity for genitivemarked objects (p < 0.0001), but not for accusative objects (p = 0.09).",
                "cite_spans": [
                    {
                        "start": 63,
                        "end": 84,
                        "text": "(Baayen et al., 2008)",
                        "ref_id": "BIBREF0"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.1.2"
            },
            {
                "text": "Although we did not find a main effect of frequency, we performed a follow-up analysis aimed to rule out the possibility that unigram frequency could be a confound for these results. Figure 1 shows that accusative-marked objects in our dataset had much higher unigram frequency in the training corpus than the genitive-marked objects. To test for the presence of the frequency effects, we re-ran the linear mixed effects analysis on surprisal scores that we normalized by subtracting the target word's log frequency from its surprisal score. The pattern remained the same: we found main effects of frequency (p = 0.006) and, as before, of case (p = 0.004), as well as an interaction between case and polarity (p < 0.0001).",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 183,
                        "end": 191,
                        "text": "Figure 1",
                        "ref_id": "FIGREF4"
                    }
                ],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.1.2"
            },
            {
                "text": "We found a main effect of case (p < 0.0001) and frequency (p = 0.001), but not of polarity (p = 0.7). There was no interaction between case and polarity (p = 0.8). Figure 4b shows there was no difference between the positive and negative conditions for either case. We observed this pattern in all experiments we ran, unless otherwise stated. ",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 164,
                        "end": 173,
                        "text": "Figure 4b",
                        "ref_id": "FIGREF12"
                    }
                ],
                "eq_spans": [],
                "section": "N-gram",
                "sec_num": null
            },
            {
                "text": "Our results suggest that the model at least learned to encode case: to predict the grammaticality pattern in (7a-7d), the model needed to infer that the grammaticality of the genitive case -but not the accusative -is constrained by the presence of negation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "5.1.3"
            },
            {
                "text": "However, these results alone are not sufficient to conclude that the model was able to infer the syntactic structure that licenses the genitive of negation. Since our experimental items had SVO word order, it could have instead learned a linear rule where the genitive-marked object is allowed whenever it follows negation. Instead, the locality constraint would predict that the object in the genitive case is licensed only when it is in the scope of negation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "5.1.3"
            },
            {
                "text": "To test whether the model has learned the locality constraint, we ran a series of experiments in which we modified our experimental sentences to include the following distractors: (1) a negated relative clause, while the genitive-marked object was licensed by the negated main clause verb, (2) a complement clause, whose polarity varied between positive and negative, and whose main clause was always negative, and (3) a negated participial construction. We give a detailed description of these constructions in the following sections.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "5.1.3"
            },
            {
                "text": "To test whether the model learned that the genitive of negation is only licensed under the scope of sentential negation, we modified the simple sentences from our dataset to include a relative clause with a negated verb. It is crucial for the model to infer the syntactic structure of these sentences: the model needs to be able to represent local scope in order to correctly predict that (8b) is ungrammatical -since the genitive-marked object in this case is outside the scope of negation. (Figure 2b) , suggesting that genitive-marked direct objects were more expected when they were licensed by the negated main clause verb. We found main effects of case (p = 0.01) and polarity (p = 0.04), and the two terms interacted (p < 0.0001). Polarity significantly affected both genitive-marked (p = 0.0001) and accusative-marked (p = 0.04) objects. Figure 3b shows that for the accusative-marked objects, the difference between positive and negative conditions was the inverse of the genitive case: an accusative-marked object was more surprising when the main clause verb was negated.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 492,
                        "end": 503,
                        "text": "(Figure 2b)",
                        "ref_id": "FIGREF6"
                    },
                    {
                        "start": 846,
                        "end": 855,
                        "text": "Figure 3b",
                        "ref_id": "FIGREF8"
                    }
                ],
                "eq_spans": [],
                "section": "Materials",
                "sec_num": "5.2.1"
            },
            {
                "text": "The analysis of frequency effects revealed that normalized surprisal scores were significantly af- fected by case (p = 0.01), frequency (p = 0.001), and the interaction of case and polarity (p < 0.0001).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Materials",
                "sec_num": "5.2.1"
            },
            {
                "text": "The trigram model's performance was the same as in Experiment 1.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "N-gram",
                "sec_num": null
            },
            {
                "text": "Our results suggest the model learned the genitivemarked object was licensed only when it appeared in the scope of negation -which in turn required the representation of syntactic structure. If the model had learned only the linear rule, it would have assigned the same surprisal in both positivegenitive and negative-genitive conditions, since both linearly followed the negation in the scope of the relative clause. The main effect of polarity suggests that the model possibly learned an interaction between case and polarity, preferring accusative objects with affirmative sentences and genitive objects under negation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "5.2.3"
            },
            {
                "text": "In the previous experiment, the distractor (i.e. the negation term that needed to be ignored) was al-ways in the relative clause. This implies that there are two possible interpretations of the results: 1) the model could represent the scope of negation and apply it to the genitive licensing rule, or 2) the model learned to ignore negation if it immediately followed the word kotoryj 'that/who', which marked the beginning of an embedded clause. To rule out the second possibility, we tested the model's performance on sentences with complement clauses. In this set of sentences, the distractor was in the main clause, while the target word (the accusative-or genitive-marked direct object) was in an embedded clause. The embedded clause varied between positive and negative polarity -and only the latter licensed the genitive object: \"The journalist didn't know that the artist's exhibition was not a failure.\"",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Materials",
                "sec_num": "5.3.1"
            },
            {
                "text": "LSTM Average surprisal was lower for genitivemarked objects when the embedded clause contained a negated verb (Figure 2c ), suggesting the model learned to represent sentential scope and did not mistake main clause negation for a licensor. Average within-item difference between positive and negative conditions was also greater for the genitive case (Figure 3c) .",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 110,
                        "end": 120,
                        "text": "(Figure 2c",
                        "ref_id": "FIGREF6"
                    },
                    {
                        "start": 351,
                        "end": 362,
                        "text": "(Figure 3c)",
                        "ref_id": "FIGREF8"
                    }
                ],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.3.2"
            },
            {
                "text": "As before, we ran a linear mixed effects model to test the significance of these findings. We found a main effect of case (p = 0.0006), as well as an interaction between case and polarity (p < 0.0001). The surprisal the language model assigned to genitive-marked objects was significantly affected by the embedded clause's polarity (p < 0.0001), while there was no such effect for the accusative case (p = 0.17).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.3.2"
            },
            {
                "text": "Our analyses of surprisal scores normalized by frequency revealed main effects of case (p = 0.0004) and frequency (0.002), as well as an interaction between case and polarity (p < 0.0001).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.3.2"
            },
            {
                "text": "N-gram The model's performance was the same as in Experiment 1.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.3.2"
            },
            {
                "text": "These results provide further evidence that the model learned the locality constraint on genitive licensing: although the main clause verb was negated in all four conditions, the surprisal the model assigned to the genitive-marked object was reduced when the verb in the embedded clause was negated as well.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "5.3.3"
            },
            {
                "text": "Experiments 2 and 3 provide some evidence that the model learned the scope constraint on the genitive of negation. However, the sentences we tested in these experiments contained overt cues that indicated the scope of negation that the model needed to ignore: in Experiment 1, the relative pronoun kotoryj indicates the beginning of the relative clause, and in Experiment 2, the pronoun chto indicates the beginning of the complement clause. Would the model be able to identify the scope of negation without these cues? We investigated this by testing the model's performance on the Russian participial construction, which has no overt function words marking the scope of negation. We constructed an experimental set of sentences which consisted of simple sentences such as those in (7a-7d) with an active present or past participle modifying the subject. \"The artist's exhibition, which did not receive attention from press, was not a failure.\"",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Materials",
                "sec_num": "5.4.1"
            },
            {
                "text": "In (10a), the genitive-marked object provala 'failure' is outside of the scope of negation, so we expected that it would be more surprising than in (10b), where the genitive is licensed by sentential scope.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Materials",
                "sec_num": "5.4.1"
            },
            {
                "text": "LSTM Figure (2d) shows the model assigned higher probability to genitive-marked objects when they were licensed by a negated verb. A linear mixed effects analysis confirmed surprisal was affected by case (p = 0.01), as well as the interaction between case and polarity (p < 0.0001). Polarity was significant for genitive-marked objects (p < 0.0001), but not for accusative-marked ones (p = 0.098).",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 5,
                        "end": 16,
                        "text": "Figure (2d)",
                        "ref_id": "FIGREF6"
                    }
                ],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.4.2"
            },
            {
                "text": "Surprisal scores normalized by frequency were significantly affected by case (p = 0.01), frequency (p = 0.003), and the interaction between polarity and case (p < 0.0001).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.4.2"
            },
            {
                "text": "N-gram The model's performance was the same as in Experiment 1.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.4.2"
            },
            {
                "text": "The model was able to capture the grammaticality pattern in (10a-10b) despite the lack of overt scope marking cues -suggesting that the model in fact represents the scope of negation instead of relying on cues such as function words introducing embedded clauses.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "5.4.3"
            },
            {
                "text": "In the experiments we have presented so far, the genitive case was always optional: genitivemarked direct objects were only grammatical in the scope of sentential negation, while the accusative case was licensed whether the sentence had positive or negative polarity. We expected to see higher surprisal for genitive-marked objects when they were outside of the scope of negation, but we did not expect any polarity-related difference for the accusative case. The situation is different in the Russian existential copula construction. First, in this construction the case alternation concerns the subject, which can be assigned the nominative or the genitive case. Second, the genitive case is always obligatory under negation. Finally, the nominative case marking is also constrained (unlike the accusative with direct objects): subjects can only receive nominative case when the sentence is affirmative. In other words, although in previous examples only the positive genitive condition was ungrammatical, in the case of the existential construction the negative nominative condition is ungrammatical as well: ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Materials",
                "sec_num": "5.5.1"
            },
            {
                "text": "LSTM A linear mixed-effects analysis revealed main effects of polarity (p < 0.0001), case (p < 0.0001), and frequency (p = 0.0003). The interaction between case and polarity was significant as well (p < 0.0001).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.5.2"
            },
            {
                "text": "We found main effects of polarity (p = 0.001), case (p = 0.0007), and frequency (p < 0.0001). There was also a significant interaction of case and polarity (p < 0.0001).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "N-gram",
                "sec_num": null
            },
            {
                "text": "The main effect of polarity shows that the model learned constraints on both the nominative and the genitive case: the genitive is licensed under negation and ungrammatical in affirmative sentences, while the opposite is true for the nominative.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "5.5.3"
            },
            {
                "text": "Further, within-item difference for both the nominative and the genitive is much bigger than in other experiments (Figure 5a ) -which suggests that the model distinguished between optionality and obligatoriness. I.e., the magnitude of surprisal was reduced in the positive-genitive condition when it was optional under negation. However, when it was required under negation, genitive-marking with positive polarity was more surprising.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 114,
                        "end": 124,
                        "text": "(Figure 5a",
                        "ref_id": "FIGREF13"
                    }
                ],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "5.5.3"
            },
            {
                "text": "Compared to previous experiments, there was a stark difference in surprisal scores between positive and negative conditions. This could be due to the fact that the the verb byt' 'to be' always appears in 3rd person singular under negation, which could have provided the model with an additional cue that the genitive case is required. 5.6 Experiment 6 5.6.1 Materials In the grammatical sentences used in Experiments 1-5, the genitive objects were directly preceded by the neg + main verb bigram, which left open the possibility that the LSTM model relied on this linear structure as a cue that the genitive case was licensed. We constructed a new dataset where the main verb was separated from the direct object by a parenthetical (e.g. \"to the surprise of the press\" in 12a-12b). If the model is learning the locality rule correctly, this parenthetical should not intervene with inferring the grammaticality pattern in 12a-12b. \"The artist's exhibition wasn't a failure, to the surprise of the press.\"",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "5.5.3"
            },
            {
                "text": "We found a main effect of case (p < 0.0004) and frequency (p = 0.01), but not of polarity (p = 0.6); there was no interaction between case and polarity (p = 0.1). Figure 4b shows there was almost no difference in surprisal the model assigned to the genitive objects licensed by negation compared to those that were ungrammatical. N-gram There was a main effect of frequency (p < 0.0001), but not of case (p = 0.34) or polarity (p = 0.96). There was no interaction between case and polarity (0.97).",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 163,
                        "end": 172,
                        "text": "Figure 4b",
                        "ref_id": "FIGREF12"
                    }
                ],
                "eq_spans": [],
                "section": "Results LSTM",
                "sec_num": "5.6.2"
            },
            {
                "text": "In (12b), the negation term was local to the target genitive object, but linearly separated from it. If the model was correctly learning the locality constraint, it would be able to predict that the genitive object provala is grammatical in (12a), but not (12b). However, the model could not identify the negation term as the licensor in these types of sentences, assigning similar surprisal to the genitive objects in (12a) and (12b). This result, however, may be due to the rarity of the parenthetical sentences in the training corpus, and does not necessarily imply the model was not learning the constraint in Experiments 1-5.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "5.6.3"
            },
            {
                "text": "In this paper, we have examined the ability of an RNN language model to learn several properties of the Russian genitive of negation. The genitive of negation can optionally mark direct objects of transitive verbs when the latter are negated, and is obligatory with subjects of existential copula constructions under negation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "General discussion and future work",
                "sec_num": "6"
            },
            {
                "text": "To be able to learn the polarity constraint on the genitive case, the model needed to represent the scope of negation. In Experiments 2 and 3, we tested this by introducing distractors to our experimental items: negated relative clauses and complement clauses that were not licensed by sentential negation. We found that the model's performance matched our predictions, assigning higher surprisal to those genitive-marked objects that were outside of the scope of negation. The results from Experiment 4 further suggest that the model could represent the scope of negation without relying on such cues as function words explicitly marking clause boundaries.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "General discussion and future work",
                "sec_num": "6"
            },
            {
                "text": "Our results from Experiment 5 provide some evidence that the model could differentiate between optionality and obligatoriness. First, we found that both the nominative and the genitive case were significantly impacted by polarity (while only the genitive was affected in other types of sentences we tested). Second, for both the nominative and the genitive case the average within-item difference between positive and negative conditions was much bigger than in other experiments. Taken together, these results suggest that the model learned that the genitive of negation was obligatory in existential sentences.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "General discussion and future work",
                "sec_num": "6"
            },
            {
                "text": "The results of Experiment 6 reveal that the model could not learn the locality constraint on the genitive of negation when the linear distance between the main verb and the direct object was increased. We tested sentences where a parenthetical intervened before the main verb and its object, and the model did not differentiate between the sentences in which the genitive object was licensed by a local negation term from those where it was not. However, this finding does not necessarily imply that the model did not learn the locality constraint in Experiments 1-5. One possible explanation for the model's behavior on the task in Experiment 6 is that constructions where a parenthetical intervenes between the main verbs and its object are not frequent in a natural corpus.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "General discussion and future work",
                "sec_num": "6"
            },
            {
                "text": "Further, more evidence is needed to asses whether the model could differentiate between syntactic structures which optionally licensed the genitive case from those where it was obligatory. One limitation of our approach is that we used the same metric for both optional and obligatory uses of the genitive of negation: we compared the surprisal the model assigned to grammatical and ungrammatical sentences, and the negated sentences with the genitive case were grammatical whether the genitive was obligatory or optional. A possible direction for future work could involve a comparison of our results to human processing data (e.g. as in . Since surprisal scores tend to correlate with reaction times (Smith and Levy, 2013), we would expect our results to match human performance.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "General discussion and future work",
                "sec_num": "6"
            },
            {
                "text": "Finally, our study only addressed some properties of the genitive of negation and only a subset of the syntactic structures in which it can appear. We haven't looked, for instance, into the genitive case marking of unaccusative subjects (13) and derived subjects of passives 14 There is also a slight difference in meaning between the genitive and accusative direct objects that we haven't addressed: while accusative direct objects usually receive a definite interpretation, the genitive ones have an existential or indefinite interpretation (Bailyn, 1997; Harves, 2002) .",
                "cite_spans": [
                    {
                        "start": 543,
                        "end": 557,
                        "text": "(Bailyn, 1997;",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 558,
                        "end": 571,
                        "text": "Harves, 2002)",
                        "ref_id": "BIBREF9"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "General discussion and future work",
                "sec_num": "6"
            },
            {
                "text": "While future investigation into these issues is needed to gain a full picture of neural network learning of the genitive of negation, our study adds to the growing body of evidence that RNN language models do not need syntactic supervision or a hierarchical bias to capture syntactic dependencies. Whether the same is true for human language learners remains to be seen.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "General discussion and future work",
                "sec_num": "6"
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Mixed-effects modeling with crossed random effects for subjects and items",
                "authors": [
                    {
                        "first": "Harald",
                        "middle": [],
                        "last": "Baayen",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Douglas",
                        "suffix": ""
                    },
                    {
                        "first": "Douglas M",
                        "middle": [],
                        "last": "Davidson",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Bates",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Journal of memory and language",
                "volume": "59",
                "issue": "4",
                "pages": "390--412",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "R Harald Baayen, Douglas J Davidson, and Douglas M Bates. 2008. Mixed-effects modeling with crossed random effects for subjects and items. Journal of memory and language, 59(4):390-412.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Existential Sentences and Negation in Russian",
                "authors": [
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Leonard",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Babby",
                        "suffix": ""
                    }
                ],
                "year": 1980,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Leonard H Babby. 1980. Existential Sentences and Negation in Russian. Karoma Publishers: Ann Ar- bor, MI.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Neural machine translation by jointly learning to align and translate",
                "authors": [
                    {
                        "first": "Dzmitry",
                        "middle": [],
                        "last": "Bahdanau",
                        "suffix": ""
                    },
                    {
                        "first": "Kyunghyun",
                        "middle": [],
                        "last": "Cho",
                        "suffix": ""
                    },
                    {
                        "first": "Yoshua",
                        "middle": [],
                        "last": "Bengio",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "International Conference for Learning Representations",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben- gio. 2014. Neural machine translation by jointly learning to align and translate. In International Con- ference for Learning Representations.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Genitive of negation is obligatory",
                "authors": [
                    {
                        "first": "F",
                        "middle": [],
                        "last": "John",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Bailyn",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "Formal approaches to slavic linguistics",
                "volume": "4",
                "issue": "",
                "pages": "84--114",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "John F Bailyn. 1997. Genitive of negation is obliga- tory. In Formal approaches to slavic linguistics, vol- ume 4, pages 84-114. University of Michigan Press: Ann Arbor, MI.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Fitting linear mixed-effects models using lme4",
                "authors": [
                    {
                        "first": "Douglas",
                        "middle": [],
                        "last": "Bates",
                        "suffix": ""
                    },
                    {
                        "first": "Martin",
                        "middle": [],
                        "last": "M\u00e4chler",
                        "suffix": ""
                    },
                    {
                        "first": "Ben",
                        "middle": [],
                        "last": "Bolker",
                        "suffix": ""
                    },
                    {
                        "first": "Steve",
                        "middle": [],
                        "last": "Walker",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1406.5823"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Douglas Bates, Martin M\u00e4chler, Ben Bolker, and Steve Walker. 2014. Fitting linear mixed-effects models using lme4. arXiv preprint arXiv:1406.5823.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Rules and representations. Behavioral and brain sciences",
                "authors": [
                    {
                        "first": "Noam",
                        "middle": [],
                        "last": "Chomsky",
                        "suffix": ""
                    }
                ],
                "year": 1980,
                "venue": "",
                "volume": "3",
                "issue": "",
                "pages": "1--15",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Noam Chomsky. 1980. Rules and representations. Be- havioral and brain sciences, 3(1):1-15.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Do RNNs learn human-like abstract word order preferences?",
                "authors": [
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Futrell",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Roger",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Levy",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1811.01866"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Richard Futrell and Roger P Levy. 2018. Do RNNs learn human-like abstract word order preferences? arXiv preprint arXiv:1811.01866.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Colorless green recurrent networks dream hierarchically",
                "authors": [
                    {
                        "first": "Kristina",
                        "middle": [],
                        "last": "Gulordava",
                        "suffix": ""
                    },
                    {
                        "first": "Piotr",
                        "middle": [],
                        "last": "Bojanowski",
                        "suffix": ""
                    },
                    {
                        "first": "Edouard",
                        "middle": [],
                        "last": "Grave",
                        "suffix": ""
                    },
                    {
                        "first": "Tal",
                        "middle": [],
                        "last": "Linzen",
                        "suffix": ""
                    },
                    {
                        "first": "Marco",
                        "middle": [],
                        "last": "Baroni",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
                "volume": "1",
                "issue": "",
                "pages": "1195--1205",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kristina Gulordava, Piotr Bojanowski, Edouard Grave, Tal Linzen, and Marco Baroni. 2018. Colorless green recurrent networks dream hierarchically. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computa- tional Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1195-1205. Associ- ation for Computational Linguistics.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "A probabilistic earley parser as a psycholinguistic model",
                "authors": [
                    {
                        "first": "John",
                        "middle": [],
                        "last": "Hale",
                        "suffix": ""
                    }
                ],
                "year": 2001,
                "venue": "Proceedings of the second meeting of the North American Chapter of the Association for Computational Linguistics on Language technologies",
                "volume": "",
                "issue": "",
                "pages": "1--8",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "John Hale. 2001. A probabilistic earley parser as a psy- cholinguistic model. In Proceedings of the second meeting of the North American Chapter of the Asso- ciation for Computational Linguistics on Language technologies, pages 1-8. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Genitive of negation and the syntax of scope",
                "authors": [
                    {
                        "first": "Stephanie",
                        "middle": [],
                        "last": "Harves",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "Proceedings of ConSOLE",
                "volume": "9",
                "issue": "",
                "pages": "96--110",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Stephanie Harves. 2002. Genitive of negation and the syntax of scope. In Proceedings of ConSOLE, vol- ume 9, pages 96-110.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Do language models understand anything? On the ability of LSTMs to understand negative polarity items",
                "authors": [
                    {
                        "first": "Jaap",
                        "middle": [],
                        "last": "Jumelet",
                        "suffix": ""
                    },
                    {
                        "first": "Dieuwke",
                        "middle": [],
                        "last": "Hupkes",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 2018 EMNLP Workshop Black-boxNLP: Analyzing and Interpreting Neural Networks for NLP",
                "volume": "",
                "issue": "",
                "pages": "222--231",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W18-5424"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jaap Jumelet and Dieuwke Hupkes. 2018. Do lan- guage models understand anything? On the ability of LSTMs to understand negative polarity items. In Proceedings of the 2018 EMNLP Workshop Black- boxNLP: Analyzing and Interpreting Neural Net- works for NLP, pages 222-231, Brussels, Belgium. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Genitive objects, existence and individuation",
                "authors": [
                    {
                        "first": "Olga",
                        "middle": [],
                        "last": "Kagan",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "",
                "volume": "34",
                "issue": "",
                "pages": "17--39",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Olga Kagan. 2010. Genitive objects, existence and in- dividuation. Russian linguistics, 34(1):17-39.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Predicting complex syntactic structure in real time: Processing of negative sentences in Russian",
                "authors": [
                    {
                        "first": "Nina",
                        "middle": [],
                        "last": "Kazanina",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "The Quarterly Journal of Experimental Psychology",
                "volume": "70",
                "issue": "11",
                "pages": "2200--2218",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Nina Kazanina. 2017. Predicting complex syntactic structure in real time: Processing of negative sen- tences in Russian. The Quarterly Journal of Experi- mental Psychology, 70(11):2200-2218.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Simple and accurate dependency parsing using bidirectional lstm feature representations",
                "authors": [
                    {
                        "first": "Eliyahu",
                        "middle": [],
                        "last": "Kiperwasser",
                        "suffix": ""
                    },
                    {
                        "first": "Yoav",
                        "middle": [],
                        "last": "Goldberg",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Transactions of the Association for Computational Linguistics",
                "volume": "4",
                "issue": "",
                "pages": "313--327",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Eliyahu Kiperwasser and Yoav Goldberg. 2016. Sim- ple and accurate dependency parsing using bidirec- tional lstm feature representations. Transactions of the Association for Computational Linguistics, 4:313-327.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Assessing the ability of LSTMs to learn syntax-sensitive dependencies",
                "authors": [
                    {
                        "first": "Tal",
                        "middle": [],
                        "last": "Linzen",
                        "suffix": ""
                    },
                    {
                        "first": "Emmanuel",
                        "middle": [],
                        "last": "Dupoux",
                        "suffix": ""
                    },
                    {
                        "first": "Yoav",
                        "middle": [],
                        "last": "Goldberg",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Transactions of the Association for Computational Linguistics",
                "volume": "4",
                "issue": "",
                "pages": "521--535",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tal Linzen, Emmanuel Dupoux, and Yoav Goldberg. 2016. Assessing the ability of LSTMs to learn syntax-sensitive dependencies. Transactions of the Association for Computational Linguistics, 4:521- 535.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Targeted syntactic evaluation of language models",
                "authors": [
                    {
                        "first": "Rebecca",
                        "middle": [],
                        "last": "Marvin",
                        "suffix": ""
                    },
                    {
                        "first": "Tal",
                        "middle": [],
                        "last": "Linzen",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "1192--1202",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Rebecca Marvin and Tal Linzen. 2018. Targeted syn- tactic evaluation of language models. In Proceed- ings of the 2018 Conference on Empirical Meth- ods in Natural Language Processing (EMNLP 2018, pages 1192-1202.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Revisiting the poverty of the stimulus: Hierarchical generalization without a hierarchical bias in recurrent neural networks",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    },
                    {
                        "first": "Thomas",
                        "middle": [],
                        "last": "Mccoy",
                        "suffix": ""
                    },
                    {
                        "first": "Robert",
                        "middle": [],
                        "last": "Frank",
                        "suffix": ""
                    },
                    {
                        "first": "Tal",
                        "middle": [],
                        "last": "Linzen",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 40th Annual Conference of the Cognitive Science Society",
                "volume": "",
                "issue": "",
                "pages": "2093--2098",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "R. Thomas McCoy, Robert Frank, and Tal Linzen. 2018. Revisiting the poverty of the stimulus: Hi- erarchical generalization without a hierarchical bias in recurrent neural networks. In Proceedings of the 40th Annual Conference of the Cognitive Science Society, pages 2093-2098, Austin, TX.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Recurrent neural network based language model",
                "authors": [
                    {
                        "first": "Tom\u00e1\u0161",
                        "middle": [],
                        "last": "Mikolov",
                        "suffix": ""
                    },
                    {
                        "first": "Martin",
                        "middle": [],
                        "last": "Karafi\u00e1t",
                        "suffix": ""
                    },
                    {
                        "first": "Luk\u00e1\u0161",
                        "middle": [],
                        "last": "Burget",
                        "suffix": ""
                    },
                    {
                        "first": "Ja\u0148",
                        "middle": [],
                        "last": "Cernock\u1ef3",
                        "suffix": ""
                    },
                    {
                        "first": "Sanjeev",
                        "middle": [],
                        "last": "Khudanpur",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Eleventh annual conference of the international speech communication association",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tom\u00e1\u0161 Mikolov, Martin Karafi\u00e1t, Luk\u00e1\u0161 Burget, Ja\u0148 Cernock\u1ef3, and Sanjeev Khudanpur. 2010. Recurrent neural network based language model. In Eleventh annual conference of the international speech com- munication association.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "The genitive subject of the verb byt' (to be)",
                "authors": [
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Elena",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Paducheva",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Studies in Polish linguistics",
                "volume": "1",
                "issue": "",
                "pages": "47--59",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Elena V Paducheva. 2004. The genitive subject of the verb byt' (to be). Studies in Polish linguistics, 1:47- 59.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "Paths and categories",
                "authors": [
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Pesetsky",
                        "suffix": ""
                    }
                ],
                "year": 1982,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "David Pesetsky. 1982. Paths and categories. Ph.D. thesis, MIT.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Can LSTM learn to capture agreement? The case of Basque",
                "authors": [
                    {
                        "first": "Shauli",
                        "middle": [],
                        "last": "Ravfogel",
                        "suffix": ""
                    },
                    {
                        "first": "Yoav",
                        "middle": [],
                        "last": "Goldberg",
                        "suffix": ""
                    },
                    {
                        "first": "Francis",
                        "middle": [],
                        "last": "Tyers",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP",
                "volume": "",
                "issue": "",
                "pages": "98--107",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Shauli Ravfogel, Yoav Goldberg, and Francis Tyers. 2018. Can LSTM learn to capture agreement? The case of Basque. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpret- ing Neural Networks for NLP, pages 98-107, Brus- sels, Belgium. Association for Computational Lin- guistics.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "The effect of word predictability on reading time is logarithmic",
                "authors": [
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Nathaniel",
                        "suffix": ""
                    },
                    {
                        "first": "Roger",
                        "middle": [],
                        "last": "Smith",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Levy",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Cognition",
                "volume": "128",
                "issue": "3",
                "pages": "302--319",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Nathaniel J Smith and Roger Levy. 2013. The effect of word predictability on reading time is logarithmic. Cognition, 128(3):302-319.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "Hierarchies in the genitive of negation",
                "authors": [
                    {
                        "first": "Alan",
                        "middle": [],
                        "last": "Timberlake",
                        "suffix": ""
                    }
                ],
                "year": 1975,
                "venue": "The Slavic and East European Journal",
                "volume": "19",
                "issue": "2",
                "pages": "123--138",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Alan Timberlake. 1975. Hierarchies in the genitive of negation. The Slavic and East European Journal, 19(2):123-138.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Grammar as a foreign language",
                "authors": [
                    {
                        "first": "Oriol",
                        "middle": [],
                        "last": "Vinyals",
                        "suffix": ""
                    },
                    {
                        "first": "\u0141ukasz",
                        "middle": [],
                        "last": "Kaiser",
                        "suffix": ""
                    },
                    {
                        "first": "Terry",
                        "middle": [],
                        "last": "Koo",
                        "suffix": ""
                    },
                    {
                        "first": "Slav",
                        "middle": [],
                        "last": "Petrov",
                        "suffix": ""
                    },
                    {
                        "first": "Ilya",
                        "middle": [],
                        "last": "Sutskever",
                        "suffix": ""
                    },
                    {
                        "first": "Geoffrey",
                        "middle": [],
                        "last": "Hinton",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Advances in neural information processing systems",
                "volume": "",
                "issue": "",
                "pages": "2773--2781",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Oriol Vinyals, \u0141ukasz Kaiser, Terry Koo, Slav Petrov, Ilya Sutskever, and Geoffrey Hinton. 2015. Gram- mar as a foreign language. In Advances in neural information processing systems, pages 2773-2781.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "What do RNN language models learn about filler-gap dependencies?",
                "authors": [
                    {
                        "first": "Ethan",
                        "middle": [],
                        "last": "Wilcox",
                        "suffix": ""
                    },
                    {
                        "first": "Roger",
                        "middle": [],
                        "last": "Levy",
                        "suffix": ""
                    },
                    {
                        "first": "Takashi",
                        "middle": [],
                        "last": "Morita",
                        "suffix": ""
                    },
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Futrell",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 2018 EMNLP Workshop Black-boxNLP: Analyzing and Interpreting Neural Networks for NLP",
                "volume": "",
                "issue": "",
                "pages": "211--221",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ethan Wilcox, Roger Levy, Takashi Morita, and Richard Futrell. 2018. What do RNN language models learn about filler-gap dependencies? In Proceedings of the 2018 EMNLP Workshop Black- boxNLP: Analyzing and Interpreting Neural Net- works for NLP, pages 211-221, Brussels, Belgium. Association for Computational Linguistics.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF1": {
                "text": "The teacher graded the homeworks.\"",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF3": {
                "text": "The artist's exhibition was a failure.\" c. negative-accusative",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF4": {
                "text": "Average unigram frequency (word count divided by the size of the training corpus) of accusative and genitive objects from our dataset.",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF5": {
                "text": "of the artist, who didn't like public attention, was not a failure.\"5.2.2 ResultsLSTM The model's surprisal was highest in the positive-genitive condition",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF6": {
                "text": "Surprisal averaged by condition (Experiments 1-4). Error bars indicate 95% confidence intervals.",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF7": {
                "text": "didn't know that the artist's exhibition was a failure.\"",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF8": {
                "text": "Within-item difference between positive and negative conditions, averaged by case (Experiments 1-4).",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF10": {
                "text": "The exhibition was not a failure.\"",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF12": {
                "text": "Surprisal averaged by condition (Experiments 5-6). Error bars indicate 95% confidence intervals.",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF13": {
                "text": "Within-item difference between positive and negative conditions, averaged by case (Experiments 5-6).",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF14": {
                "text": "The artist's exhibition was a failure, to the surprise of the press.\"",
                "type_str": "figure",
                "num": null,
                "uris": null
            }
        }
    }
}