Base_JSON/prefixD/json/D16/D16-1009.json

{
    "paper_id": "D16-1009",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T16:35:52.531176Z"
    },
    "title": "Modeling Human Reading with Neural Attention",
    "authors": [
        {
            "first": "Michael",
            "middle": [],
            "last": "Hahn",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Edinburgh",
                "location": {
                    "addrLine": "10 Crichton Street",
                    "postCode": "EH8 9AB",
                    "settlement": "Edinburgh",
                    "country": "UK"
                }
            },
            "email": ""
        },
        {
            "first": "Frank",
            "middle": [],
            "last": "Keller",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Edinburgh",
                "location": {
                    "addrLine": "10 Crichton Street",
                    "postCode": "EH8 9AB",
                    "settlement": "Edinburgh",
                    "country": "UK"
                }
            },
            "email": "keller@inf.ed.ac.uk"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "When humans read text, they fixate some words and skip others. However, there have been few attempts to explain skipping behavior with computational models, as most existing work has focused on predicting reading times (e.g., using surprisal). In this paper, we propose a novel approach that models both skipping and reading, using an unsupervised architecture that combines a neural attention with autoencoding, trained on raw text using reinforcement learning. Our model explains human reading behavior as a tradeoff between precision of language understanding (encoding the input accurately) and economy of attention (fixating as few words as possible). We evaluate the model on the Dundee eye-tracking corpus, showing that it accurately predicts skipping behavior and reading times, is competitive with surprisal, and captures known qualitative features of human reading.",
    "pdf_parse": {
        "paper_id": "D16-1009",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "When humans read text, they fixate some words and skip others. However, there have been few attempts to explain skipping behavior with computational models, as most existing work has focused on predicting reading times (e.g., using surprisal). In this paper, we propose a novel approach that models both skipping and reading, using an unsupervised architecture that combines a neural attention with autoencoding, trained on raw text using reinforcement learning. Our model explains human reading behavior as a tradeoff between precision of language understanding (encoding the input accurately) and economy of attention (fixating as few words as possible). We evaluate the model on the Dundee eye-tracking corpus, showing that it accurately predicts skipping behavior and reading times, is competitive with surprisal, and captures known qualitative features of human reading.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Humans read text by making a sequence of fixations and saccades. During a fixation, the eyes land on a word and remain fairly static for 200-250 ms. Saccades are the rapid jumps that occur between fixations, typically lasting 20-40 ms and spanning 7-9 characters (Rayner, 1998) . Readers, however, do not simply fixate one word after another; some saccades go in reverse direction, and some words are fixated more than once or skipped altogether.",
                "cite_spans": [
                    {
                        "start": 263,
                        "end": 277,
                        "text": "(Rayner, 1998)",
                        "ref_id": "BIBREF24"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "A range of computational models have been developed to account for human eye-movements in reading (Rayner and Reichle, 2010) , including models of saccade generation in cognitive psychology, such as EZ-Reader (Reichle et al., 1998 (Reichle et al., , 2003 (Reichle et al., , 2009 , SWIFT (Engbert et al., 2002 (Engbert et al., , 2005 , or the Bayesian Model of Bicknell and Levy (2010) . More recent approaches use machine learning models trained on eye-tracking data to predict human reading patterns Nivre, 2009, 2010; Hara et al., 2012; Matthies and S\u00f8gaard, 2013) . Both types of models involve theoretical assumptions about human eye-movements, or at least require the selection of relevant eye-movement features. Model parameters have to be estimated in a supervised way from eye-tracking corpora.",
                "cite_spans": [
                    {
                        "start": 98,
                        "end": 124,
                        "text": "(Rayner and Reichle, 2010)",
                        "ref_id": "BIBREF26"
                    },
                    {
                        "start": 209,
                        "end": 230,
                        "text": "(Reichle et al., 1998",
                        "ref_id": "BIBREF27"
                    },
                    {
                        "start": 231,
                        "end": 254,
                        "text": "(Reichle et al., , 2003",
                        "ref_id": "BIBREF29"
                    },
                    {
                        "start": 255,
                        "end": 278,
                        "text": "(Reichle et al., , 2009",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 287,
                        "end": 308,
                        "text": "(Engbert et al., 2002",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 309,
                        "end": 332,
                        "text": "(Engbert et al., , 2005",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 360,
                        "end": 384,
                        "text": "Bicknell and Levy (2010)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 501,
                        "end": 519,
                        "text": "Nivre, 2009, 2010;",
                        "ref_id": null
                    },
                    {
                        "start": 520,
                        "end": 538,
                        "text": "Hara et al., 2012;",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 539,
                        "end": 566,
                        "text": "Matthies and S\u00f8gaard, 2013)",
                        "ref_id": "BIBREF16"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Unsupervised approaches, that do not involve training the model on eye-tracking data, have also been proposed. A key example is surprisal, which measures the predictability of a word in context, defined as the negative logarithm of the conditional probability of the current word given the preceding words (Hale, 2001; Levy, 2008) . Surprisal is computed by a language model, which can take the form of a probabilistic grammar, an n-gram model, or a recurrent neural network. While surprisal has been shown to correlate with word-by-word reading times (McDonald and Shillcock, 2003a,b; Demberg and Keller, 2008; Frank and Bod, 2011; Smith and Levy, 2013) , it cannot explain other aspects of human reading, such as reverse saccades, re-fixations, or skipping. Skipping is a particularly intriguing phenomenon: about 40% of all words are skipped (in the Dundee corpus, see below), without apparent detriment to text understanding.",
                "cite_spans": [
                    {
                        "start": 306,
                        "end": 318,
                        "text": "(Hale, 2001;",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 319,
                        "end": 330,
                        "text": "Levy, 2008)",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 552,
                        "end": 585,
                        "text": "(McDonald and Shillcock, 2003a,b;",
                        "ref_id": null
                    },
                    {
                        "start": 586,
                        "end": 611,
                        "text": "Demberg and Keller, 2008;",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 612,
                        "end": 632,
                        "text": "Frank and Bod, 2011;",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 633,
                        "end": 654,
                        "text": "Smith and Levy, 2013)",
                        "ref_id": "BIBREF30"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In this paper, we propose a novel model architecture that is able to explain which words are skipped and which ones are fixated, while also predicting reading times for fixated words. Our approach is completely unsupervised and requires only unlabeled text for training.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Compared to language as a whole, reading is a recent innovation in evolutionary terms, and people learning to read do not have access to competent readers' eye-movement patterns as training data. This suggests that human eye-movement patterns emerge from general principles of language processing that are independent of reading. Our starting point is the Tradeoff Hypothesis: Human reading optimizes a tradeoff between precision of language understanding (encoding the input accurately) and economy of attention (fixating as few words as possible). Based on the Tradeoff Hypothesis, we expect that humans only fixate words to the extent necessary for language understanding, while skipping words whose contribution to the overall meaning can be inferred from context.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In order to test these assumptions, this paper investigates the following questions:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "1. Can the Tradeoff Hypothesis be implemented in an unsupervised model that predicts skipping and reading times in quantitative terms? In particular, can we compute surprisal based only on the words that are actually fixated?",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "2. Can the Tradeoff Hypothesis explain known qualitative features of human fixation patterns? These include dependence on word frequency, word length, predictability in context, a contrast between content and function words, and the statistical dependence of the current fixation on previous fixations.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "To investigate these questions, we develop a generic architecture that combines neural language modeling with recent ideas on integrating recurrent neural networks with mechanisms of attention, which have shown promise both in NLP and in computer vision. We train our model end-to-end on a large text corpus to optimize a tradeoff between minimizing input reconstruction error and minimizing the number of words fixated. We evaluate the model's reading behavior against a corpus of human eye-tracking data. Apart from the unlabeled training corpus and the generic architecture, no further assumptions about language structure are made -in particular, no lex-icon or grammar or otherwise labeled data is required. Our unsupervised model is able to predict human skips and fixations with an accuracy of 63.7%. This compares to a baseline of 52.6% and a supervised accuracy of 69.9%. For fixated words, the model significantly predicts human reading times in a linear mixed effects analysis. The performance of our model is comparable to surprisal, even though it only fixates 60.4% of all input words. Furthermore, we show that known qualitative features of human fixation sequences emerge in our model without additional assumptions.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "A range of attention-based neural network architectures have recently been proposed in the literature, showing promise in both NLP and computer vision (e.g., Mnih et al., 2014; Bahdanau et al., 2015) . Such architectures incorporate a mechanism that allows the network to dynamically focus on a restricted part of the input. Attention is also a central concept in cognitive science, where it denotes the focus of cognitive processing. In both language processing and visual processing, attention is known to be limited to a restricted area of the visual field, and shifts rapidly through eye-movements (Henderson, 2003) .",
                "cite_spans": [
                    {
                        "start": 158,
                        "end": 176,
                        "text": "Mnih et al., 2014;",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 177,
                        "end": 199,
                        "text": "Bahdanau et al., 2015)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 602,
                        "end": 619,
                        "text": "(Henderson, 2003)",
                        "ref_id": "BIBREF11"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "Attention-based neural architectures either employ soft attention or hard attention. Soft attention distributes real-valued attention values over the input, making end-to-end training with gradient descent possible. Hard attention mechanisms make discrete choices about which parts of the input to focus on, and can be trained with reinforcement learning (Mnih et al., 2014) . In NLP, soft attention can mitigate the difficulty of compressing long sequences into fixed-dimensional vectors, with applications in machine translation (Bahdanau et al., 2015) and question answering (Hermann et al., 2015) . In computer vision, both types of attention can be used for selecting regions in an image (Ba et al., 2015; Xu et al., 2015) .",
                "cite_spans": [
                    {
                        "start": 355,
                        "end": 374,
                        "text": "(Mnih et al., 2014)",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 531,
                        "end": 554,
                        "text": "(Bahdanau et al., 2015)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 578,
                        "end": 600,
                        "text": "(Hermann et al., 2015)",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 693,
                        "end": 710,
                        "text": "(Ba et al., 2015;",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 711,
                        "end": 727,
                        "text": "Xu et al., 2015)",
                        "ref_id": "BIBREF34"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "ing and economy of attention. We make this idea explicit by proposing NEAT (NEural Attention Tradeoff), a model that reads text and attempts to reconstruct it afterwards. While reading, the network chooses which words to process and which words to skip. The Tradeoff Hypothesis is formalized using a training objective that combines accuracy of reconstruction with economy of attention, encouraging the network to only look at words to the extent that is necessary for reconstructing the sentence.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "We use a neural sequence-to-sequence architecture (Sutskever et al., 2014) with a hard attention mechanism. We illustrate the model in Figure 1 , operating on a three-word sequence w w w. The most basic components are the reader, labeled R, and the decoder. Both of them are recurrent neural networks with Long Short-Term Memory (LSTM, Hochreiter and Schmidhuber, 1997) units. The recurrent reader network is expanded into time steps R 0 , . . . , R 3 in the figure. It goes over the input sequence, reading one word w i at a time, and converts the word sequence into a sequence of vectors h 0 , . . . , h 3 . Each vector h i acts as a fixed-dimensionality encoding of the word sequence w 1 , . . . , w i that has been read so far. The last vector h 3 (more generally h N for sequence length N), which encodes the entire input sequence, is then fed into the input layer of the decoder network, which attempts to reconstruct the input sequence w w w. It is also realized as a recurrent neural network, collapsed into a single box in the figure. It models a probability distribution over word sequences, outputting a probability distribution P Decoder (w i |w w w 1,...,i\u22121 , h N ) over the vocabulary in the i-th step, as is common in neural language modeling (Mikolov et al., 2010) . As the decoder has access to the vector representation created by the reader network, it ideally is able to assign the highest probability to the word sequence w w w that was actually read. Up to this point, the model is a standard sequence-to-sequence architecture reconstructing the input sequence, that is, performing autoencoding.",
                "cite_spans": [
                    {
                        "start": 50,
                        "end": 74,
                        "text": "(Sutskever et al., 2014)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 336,
                        "end": 369,
                        "text": "Hochreiter and Schmidhuber, 1997)",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 1259,
                        "end": 1281,
                        "text": "(Mikolov et al., 2010)",
                        "ref_id": "BIBREF19"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 135,
                        "end": 143,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Architecture",
                "sec_num": "3.1"
            },
            {
                "text": "As a basic model of human processing, NEAT contains two further components. First, experimental evidence shows that during reading, humans constantly make predictions about the upcoming input (e.g., Van Gompel and Pickering, 2007) . As a model of this behavior, the reader network at each time step outputs a probability distribution P R over the lexicon. This distribution describes which words are likely to come next (i.e., the reader network performs language modeling). Unlike the modeling performed by the decoder, P R , via its recurrent connections, has access to the previous context only. Second, we model skipping by stipulating that only some of the input words w i are fed into the reader network R, while R receives a special vector representation, containing no information about the input word, in other cases. These are the words that are skipped. In NEAT, at each time step during reading, the attention module A decides whether the next word is shown to the reader network or not. When humans skip a word, they are able to identify it using parafoveal preview (Rayner, 2009) . Thus, we can assume that the choice of which words to skip takes into account not only the prior context but also a preview of the word itself. We therefore allow the attention module to take the input word into account when making its decision. In addition, the attention module has access to the previous state h i\u22121 of the reader network, which summarizes what has been read so far. To allow for interaction between skipping and prediction, we also give the attention module access to the probability of the input word according to the prediction P R made at the last time step. If we write the decision made by A as \u03c9 i \u2208 {0, 1}, where \u03c9 i = 1 means that word w i is shown to the reader and 0 means that it is not, we can write the probability of showing word w i as:",
                "cite_spans": [
                    {
                        "start": 203,
                        "end": 230,
                        "text": "Gompel and Pickering, 2007)",
                        "ref_id": "BIBREF32"
                    },
                    {
                        "start": 1079,
                        "end": 1093,
                        "text": "(Rayner, 2009)",
                        "ref_id": "BIBREF25"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Architecture",
                "sec_num": "3.1"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "P(\u03c9 i = 1|\u03c9 \u03c9 \u03c9 1...i\u22121 , w w w) = P A (w i , h i\u22121 , P R (w i |w w w 1...i\u22121 , \u03c9 \u03c9 \u03c9 1...i\u22121 ))",
                        "eq_num": "(1)"
                    }
                ],
                "section": "Architecture",
                "sec_num": "3.1"
            },
            {
                "text": "We implement A as a feed-forward network, followed by taking a binary sample \u03c9 i . We obtain the surprisal of an input word by taking the negative logarithm of the conditional probability of this word given the context words that precede it:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Architecture",
                "sec_num": "3.1"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "Surp(w i |w w w 1...i\u22121 ) = \u2212 log P R (w i |w w w 1...i\u22121 , \u03c9 \u03c9 \u03c9 1...i\u22121 )",
                        "eq_num": "(2)"
                    }
                ],
                "section": "Architecture",
                "sec_num": "3.1"
            },
            {
                "text": "As a consequence of skipping, not all input words are accessible to the reader network. Therefore, the probability and surprisal estimates it computes crucially only take into account the words that have actually been fixated. We will refer to this quantity as the restricted surprisal, as opposed to full surprisal, which is computed based on all prior context words. The key quantities for predicting human reading are the fixation probabilities in equation 1, which model fixations and skips, and restricted surprisal in equation 2, which models the reading times of the words that are fixated.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Architecture",
                "sec_num": "3.1"
            },
            {
                "text": "Given network parameters \u03b8 and a sequence w w w of words, the network stochastically chooses a sequence \u03c9 \u03c9 \u03c9 according to (1) and incurs a loss L(\u03c9 \u03c9 \u03c9|w w w, \u03b8) for language modeling and reconstruction:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Model Objective",
                "sec_num": "3.2"
            },
            {
                "text": "L(\u03c9 \u03c9 \u03c9|w w w, \u03b8) = \u2212 \u2211 i log P R (w i |w w w 1,...,i\u22121 , \u03c9 \u03c9 \u03c9 1,...,i\u22121 ; \u03b8) \u2212 \u2211 i log P Decoder (w i |w w w 1,...,i\u22121 ; h N ; \u03b8) (3)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Model Objective",
                "sec_num": "3.2"
            },
            {
                "text": "where P R (w i , . . . ) denotes the output of the reader after reading w i\u22121 , and P Decoder (w i | . . . ; h N ) is the output of the decoder at time i \u2212 1, with h N being the vector representation created by the reader network for the entire input sequence.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Model Objective",
                "sec_num": "3.2"
            },
            {
                "text": "To implement the Tradeoff Hypothesis, we train NEAT to solve language modeling and reconstruction with minimal attention, i.e., the network minimizes the expected loss:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Model Objective",
                "sec_num": "3.2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "Q(\u03b8) := E w w w,\u03c9 \u03c9 \u03c9 [L(\u03c9 \u03c9 \u03c9|w w w, \u03b8) + \u03b1 \u2022 \u03c9 \u03c9 \u03c9 1 ]",
                        "eq_num": "(4)"
                    }
                ],
                "section": "Model Objective",
                "sec_num": "3.2"
            },
            {
                "text": "where word sequences w w w are drawn from a corpus, and \u03c9 \u03c9 \u03c9 is distributed according to P(\u03c9 \u03c9 \u03c9|w w w, \u03b8) as defined in (1). In (4), \u03c9 \u03c9 \u03c9 1 is the number of words shown to the reader, and \u03b1 > 0 is a hyperparameter. The term \u03b1 \u2022 \u03c9 \u03c9 \u03c9 1 encourages NEAT to attend to as few words as possible. Note that we make no assumption about linguistic structure -the only ingredients of NEAT are the neural architecture, the objective (4), and the corpus from which the sequences w w w are drawn.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Model Objective",
                "sec_num": "3.2"
            },
            {
                "text": "We follow previous approaches to hard attention in using a combination of gradient descent and reinforcement learning, and separate the training of the recurrent networks from the training of A. To train the reader R and the decoder, we temporarily remove the attention network A, set \u03c9 \u03c9 \u03c9 \u223c Binom(n, p) (n sequence length, p a hyperparameter), and minimize E[L(w w w|\u03b8, \u03c9 \u03c9 \u03c9)] using stochastic gradient descent, sampling a sequence \u03c9 \u03c9 \u03c9 for each input sequence. In effect, NEAT is trained to perform reconstruction and language modeling when there is noise in the input. After R and the decoder have been trained, we fix their parameters and train A using the RE-INFORCE rule (Williams, 1992) , which performs stochastic gradient descent using the estimate",
                "cite_spans": [
                    {
                        "start": 680,
                        "end": 696,
                        "text": "(Williams, 1992)",
                        "ref_id": "BIBREF33"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training",
                "sec_num": "3.3"
            },
            {
                "text": "1 |B| \u2211 w w w\u2208B;\u03c9 \u03c9 \u03c9 (L(\u03c9 \u03c9 \u03c9|w w w, \u03b8) + \u03b1 \u2022 \u03c9 \u03c9 \u03c9 1 ) \u2202 \u03b8 A (log P(\u03c9 \u03c9 \u03c9|w w w, \u03b8)) (5) for the gradient \u2202 \u03b8 A Q.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training",
                "sec_num": "3.3"
            },
            {
                "text": "Here, B is a minibatch, \u03c9 \u03c9 \u03c9 is sampled from P(\u03c9 \u03c9 \u03c9|w w w, \u03b8), and \u03b8 A \u2282 \u03b8 is the set of parameters of A. For reducing the variance of this estimator, we subtract in the i-th step an estimate of the expected loss:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training",
                "sec_num": "3.3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "U(w w w, \u03c9 \u03c9 \u03c9 1...i\u22121 ) := E \u03c9 \u03c9 \u03c9 i...N [L(\u03c9 \u03c9 \u03c9 1...i\u22121 \u03c9 \u03c9 \u03c9 i...N |w w w, \u03b8) + \u03b1 \u2022 \u03c9 \u03c9 \u03c9 1 ]",
                        "eq_num": "(6)"
                    }
                ],
                "section": "Training",
                "sec_num": "3.3"
            },
            {
                "text": "We compute the expected loss using an LSTM that we train simultaneously with A to predict L + \u03b1 \u03c9 \u03c9 \u03c9 1 based on w w w and \u03c9 \u03c9 \u03c9 1...i\u22121 . To make learning more stable, we add an entropy term encouraging the distribution to be smooth, following Xu et al. (2015) . The parameter updates to A are thus:",
                "cite_spans": [
                    {
                        "start": 245,
                        "end": 261,
                        "text": "Xu et al. (2015)",
                        "ref_id": "BIBREF34"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training",
                "sec_num": "3.3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "\u2211 w w w,\u03c9 \u03c9 \u03c9 \u2211 i (L(\u03c9 \u03c9 \u03c9|w w w, \u03b8) + \u03b1 \u03c9 \u03c9 \u03c9 1 \u2212U(w w w, \u03c9 \u03c9 \u03c9 1...i\u22121 )) \u2022 \u2202 \u03b8 A (log P(\u03c9 i |\u03c9 \u03c9 \u03c9 1...i\u22121 , w w w, \u03b8)) \u2212\u03b3 \u2202 \u03b8 A \u2211 w w w,\u03c9 \u03c9 \u03c9 \u2211 i H[P(\u03c9 i |\u03c9 \u03c9 \u03c9 1,...,i\u22121 , w w w, \u03b8)]",
                        "eq_num": "(7)"
                    }
                ],
                "section": "Training",
                "sec_num": "3.3"
            },
            {
                "text": "where \u03b3 is a hyperparameter, and H the entropy.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training",
                "sec_num": "3.3"
            },
            {
                "text": "Our aim is to evaluate how well NEAT predicts human fixation behavior and reading times. Furthermore, we want show that known qualitative properties emerge from the Tradeoff Hypothesis, even though no prior knowledge about useful features is hard-wired in NEAT.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methods",
                "sec_num": "4"
            },
            {
                "text": "w 1 w 2 w 3 A A A R 0 R 1 R 2 R 3 Decoder h 0 h 1 h 2 P R 1 P R 2 P R 3",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methods",
                "sec_num": "4"
            },
            {
                "text": "Figure 1: The architecture of the proposed model, reading a three-word input sequence w 1 , w 2 , w 3 . R is the reader network and P R the probability distribution it computes in each time step. A is the attention network. At each time step, the input, its probability according to P R , and the previous state h i\u22121 of R are fed into A, which then decides whether the word is read or skipped.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methods",
                "sec_num": "4"
            },
            {
                "text": "For both the reader and the decoder networks, we choose a one-layer LSTM network with 1,000 memory cells. The attention network is a one-layer feedforward network. For the loss estimator U, we use a bidirectional LSTM with 20 memory cells. Input data is split into sequences of 50 tokens, which are used as the input sequences for NEAT, disregarding sentence boundaries. Word embeddings have 100 dimensions, are shared between the reader and the attention network, and are only trained during the training of the reader. The vocabulary consists of the 10,000 most frequent words from the training corpus. We trained NEAT on the training set of the Daily Mail section of the corpus described by Hermann et al. (2015), which consists of 195,462 articles from the Daily Mail newspaper, containing approximately 200 million tokens. The recurrent networks and the attention network were each trained for one epoch. For initialization, weights are drawn from the uniform distribution. We set \u03b1 = 5.0, \u03b3 = 5.0, and used a constant learning rate of 0.01 for A.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training Setup",
                "sec_num": "4.1"
            },
            {
                "text": "For evaluation, we use the English section of the Dundee corpus (Kennedy and Pynte, 2005) , which consists of 20 texts from The Independent, annotated with eye-movement data from ten English native speakers. Each native speakers read all 20 texts and answered a comprehension question after each text. We split the Dundee corpus into a development and a test set, with texts 1-3 constituting the development set. The development set consists of 78,300 tokens, and the test set of 281,911 tokens. For evaluation, we removed the datapoints removed by Demberg and Keller (2008) , mainly consisting of words at the beginning or end of lines, outliers, and cases of track loss. Furthermore, we removed datapoints where the word was outside of the vocabulary of the model, and those datapoints mapped to positions 1-3 or 48-50 of a sequence when splitting the data. After preprocessing, 62.9% of the development tokens and 64.7% of the test tokens remained. To obtain the number of fixations on a token and reading times, we used the eye-tracking measures computed by Demberg and Keller (2008) . The overall fixation rate was 62.1% on the development set, and 61.3% on the test set. The development set was used to run preliminary versions of the human evaluation studies, and to determine the human skipping rate (see Section 5). All the results reported in this paper were computed on the test set, which remained unseen until the model was final.",
                "cite_spans": [
                    {
                        "start": 64,
                        "end": 89,
                        "text": "(Kennedy and Pynte, 2005)",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 549,
                        "end": 574,
                        "text": "Demberg and Keller (2008)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 1062,
                        "end": 1087,
                        "text": "Demberg and Keller (2008)",
                        "ref_id": "BIBREF5"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Corpus",
                "sec_num": "4.2"
            },
            {
                "text": "Throughout this section, we consider the following baselines for the attention network: random attention is defined by \u03c9 \u03c9 \u03c9 \u223c Binom(n, p), with p = 0.62, the human fixation rate in the development set. For full attention, we take \u03c9 \u03c9 \u03c9 = 1, i.e., all words are fixated. We also derive fixation predictions from full surprisal, word frequency, and word length by choosing a threshold such that the resulting fixation rate matches the human fixation rate on the development set.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results and Discussion",
                "sec_num": "5"
            },
            {
                "text": "By averaging over all possible fixation sequences, NEAT defines for each word in a sequence a probability that it will be fixated. This probability is not efficiently computable, so we approximate it by sampling a sequence \u03c9 \u03c9 \u03c9 and taking the probabilities P(\u03c9 i = 1|\u03c9 1...i\u22121 , w w w) for i = 1, . . . , 50. These sim-ulated fixation probabilities can be interpreted as defining a distribution of attention over the input sequence. Figure 2 shows heatmaps of the simulated and human fixation probabilities, respectively, for the beginning of a text from the Dundee corpus. While some differences between simulated and human fixation probabilities can be noticed, there are similarities in the general qualitative features of the two heatmaps. In particular, function words and short words are less likely to be fixated than content words and longer words in both the simulated and the human data.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 434,
                        "end": 442,
                        "text": "Figure 2",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Quantitative Properties",
                "sec_num": "5.1"
            },
            {
                "text": "We first evaluate NEAT intrinsically by measuring how successful the network is at predicting the next word and reconstructing the input while minimizing the number of fixations. We compare perplexity on reconstruction and language modeling for \u03c9 \u03c9 \u03c9 \u223c P(\u03c9 \u03c9 \u03c9|w w w, \u03b8). In addition to the baselines, we run NEAT on the fixations generated by the human readers of the Dundee corpus, i.e., we use the human fixation sequence as \u03c9 \u03c9 \u03c9 instead of the fixation sequence generated by A to compute perplexity. This will tell us to what extent the human behavior minimizes the NEAT objective (4).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Reconstruction and Language Modeling",
                "sec_num": null
            },
            {
                "text": "The results are given in Table 1 . In all settings, the fixation rates are similar (60.4% to 62.1%) which makes the perplexity figures directly comparable. While NEAT has a higher perplexity on both tasks compared to full attention, it considerably outperforms random attention. It also outperforms the word length, word frequency, and full surprisal baselines. The perplexity on human fixation sequences is similar to that achieved using word frequency. Based on these results, we conclude that REINFORCE successfully optimizes the objective (4).",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 25,
                        "end": 32,
                        "text": "Table 1",
                        "ref_id": "TABREF3"
                    }
                ],
                "eq_spans": [],
                "section": "Reconstruction and Language Modeling",
                "sec_num": null
            },
            {
                "text": "Human reading behavior is stochastic in the sense that different runs of eye-tracking experiments such as the ones recorded in the Dundee corpus yield different eye-movement sequences. NEAT is also stochastic, in the sense that, given a word sequence w, it defines a probability distribution over fixation sequences \u03c9 \u03c9 \u03c9. Ideally, this distribution should be close to the actual distribution of fixation sequences produced by humans reading the sequence, as measured by perplexity.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Likelihood of Fixation Data",
                "sec_num": null
            },
            {
                "text": "We find that the perplexity of the fixation se-Acc F1 fix F1 skip NEAT 63.7 70.4 53.0 Supervised Models Nilsson and Nivre (2009) 69.5 75.2 62.6 Matthies and S\u00f8gaard (2013) quences produced by the ten readers in the Dundee corpus under NEAT is 1.84. A perplexity of 2.0 corresponds to the random baseline Binom(n, 0.5), and a perplexity of 1.96 to random attention Binom(n, 0.62). As a lower bound on what can achieved with models disregarding the context, using the human fixation rates for each word as probabilities, we obtain a perplexity of 1.68.",
                "cite_spans": [
                    {
                        "start": 104,
                        "end": 128,
                        "text": "Nilsson and Nivre (2009)",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 144,
                        "end": 171,
                        "text": "Matthies and S\u00f8gaard (2013)",
                        "ref_id": "BIBREF16"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Likelihood of Fixation Data",
                "sec_num": null
            },
            {
                "text": "Previous work on supervised models for modeling fixations (Nilsson and Nivre, 2009; Matthies and S\u00f8gaard, 2013) has been evaluated by measuring the overlap of the fixation sequences produced by the models with those in the Dundee corpus. For NEAT, this method of evaluation is problematic as differences between model predictions and human data may be due to differences in the rate of skipping, and due to the inherently stochastic nature of fixations. We therefore derive model predictions by rescaling the simulated fixation probabilities so that their average equals the fixation rate in the development set, and then greedily take the maximum-likelihood sequence. That is, we predict a fixation if the rescaled probability is greater than 0.5, and a skip otherwise. As in previous work, we report the accuracy of fixations and skips, and also separate F1 scores for fixations and skips. As lower and upper bounds, we use the random baseline \u03c9 \u03c9 \u03c9 \u223c Binom(n, 0.62) and the agreement of the ten human readers, respectively. The re-",
                "cite_spans": [
                    {
                        "start": 58,
                        "end": 83,
                        "text": "(Nilsson and Nivre, 2009;",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 84,
                        "end": 111,
                        "text": "Matthies and S\u00f8gaard, 2013)",
                        "ref_id": "BIBREF16"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Accuracy of Fixation Sequences",
                "sec_num": null
            },
            {
                "text": "The decision of the Human Fertility and Embryology Authority (HFEA) to allow a couple to select genetically their next baby was bound to raise concerns that advances in biotechnology are racing ahead of our ability to control the consequences. The couple at the centre of this case have a son who suffers from a potentially fatal disorder and whose best hope is a marrow transplant from a sibling, so the stakes of this decision are particularly high. The HFEA's critics believe that it sanctions 'designer babies' and does not show respect for the sanctity of individual life. Certainly, the authority's backing for Shahana and Raj Hashmi's plea for genetic screening raises fundamental questions",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Accuracy of Fixation Sequences",
                "sec_num": null
            },
            {
                "text": "The decision of the Human Fertility and Embryology Authority (HFEA) to allow a couple to select genetically their next baby was bound to raise concerns that advances in biotechnology are racing ahead of our ability to control the consequences. The couple at the centre of this case have a son who suffers from a potentially fatal disorder and whose best hope is a marrow transplant from a sibling, so the stakes of this decision are particularly high. The HFEA's critics believe that it sanctions 'designer babies' and does not show respect for the sanctity of individual life. Certainly, the authority's backing for Shahana and Raj Hashmi's plea for genetic screening raises fundamental questions sults are shown in Table 2 . NEAT clearly outperforms the random baseline and shows results close to full surprisal (where we apply the same rescaling and thresholding as for NEAT). This is remarkable given that NEAT has access to only 60.4% of the words in the corpus in order to predict skipping, while full surprisal has access to all the words. Word frequency and word length perform well, almost reaching the performance of supervised models. This shows that the bulk of skipping behavior is already explained by word frequency and word length effects. Note, however, that NEAT is completely unsupervised, and does not know that it has to pay attention to word frequency; this is something the model is able to infer.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 717,
                        "end": 724,
                        "text": "Table 2",
                        "ref_id": "TABREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Accuracy of Fixation Sequences",
                "sec_num": null
            },
            {
                "text": "To evaluate the predictions NEAT makes for reading times, we use linear mixed-effects models containing restricted surprisal derived from NEAT for the Dundee test set. The mixed models also include a set of standard baseline predictors, viz., word length, log word frequency, log frequency of the previous word, launch distance, landing position, and the position of the word in the sentence. We treat participants and items as random factors. As the dependent variable, we take first pass duration, which is the sum of the durations of all fixations from first entering the word to first leaving it. We compare against full surprisal as an upper bound and against random surprisal as a lower bound. Random surprisal is surprisal computed by a model with random attention; this allows us to assess how much surprisal degrades when only 60.4% of all words are fixated, but no information is available as to which words should be fixated. The results in Table 3 show that restricted surprisal as computed by NEAT, full surprisal, and random surprisal are all significant predictors of reading time.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 952,
                        "end": 959,
                        "text": "Table 3",
                        "ref_id": "TABREF5"
                    }
                ],
                "eq_spans": [],
                "section": "Restricted Surprisal and Reading Times",
                "sec_num": null
            },
            {
                "text": "In order to compare the three surprisal estimates, we therefore need a measure of effect size. For this, we compare the model fit of the three mixed effects models using deviance, which is defined as the difference between the log likelihood of the model under consideration minus the log likelihood of the baseline model, multiplied by \u22122. Higher de- viance indicates greater improvement in model fit over the baseline model. We find that the mixed model that includes restricted surprisal achieves a deviance of 867, compared to the model containing only the baseline features. With full surprisal, we obtain a deviance of 980. On the other hand, the model including random surprisal achieves a lower deviance of 832. This shows that restricted surprisal as computed by NEAT not only significantly predicts reading times, it also provides an improvement in model fit compared to the baseline predictors. Such an improvement is also observed with random surprisal, but restricted surprisal achieves a greater improvement in model fit. Full surprisal achieves an even greater improvement, but this is not unexpected, as full surprisal has access to all words, unlike NEAT or random surprisal, which only have access to 60.4% of the words.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Restricted Surprisal and Reading Times",
                "sec_num": null
            },
            {
                "text": "We now examine the second key question we defined in Section 1, investigating the qualitative features of the simulated fixation sequences. We will focus on comparing the predictions of NEAT with that of word frequency, which performs comparably at the task of predicting fixation sequences (see Sec- tion 5.1). We show NEAT nevertheless makes relevant predictions that go beyond frequency.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative Properties",
                "sec_num": "5.2"
            },
            {
                "text": "Fixations of Successive Words While predictors derived from word frequency treat the decision whether to fixate or skip words as independent, humans are more likely to fixate a word when the previous word was skipped (Rayner, 1998) . This effect is also seen in NEAT. More precisely, both in the human data and in the simulated fixation data, the conditional fixation probability P(\u03c9 i = 1|\u03c9 i\u22121 = 1) is lower than the marginal probability P(\u03c9 i = 1). The ratio of these probabilities is 0.85 in the human data, and 0.81 in NEAT. The threshold predictor derived from word frequency also shows this effect (as the frequencies of successive words are not independent), but it is weaker (ratio 0.91).",
                "cite_spans": [
                    {
                        "start": 217,
                        "end": 231,
                        "text": "(Rayner, 1998)",
                        "ref_id": "BIBREF24"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative Properties",
                "sec_num": "5.2"
            },
            {
                "text": "To further test the context dependence of NEAT's fixation behavior, we ran a mixed model predicting the fixation probabilities simulated by NEAT, with items as random factor and the log frequency of word i as predictor. Adding \u03c9 i\u22121 as a predictor results in a significant improvement in model fit (deviance = 4,798, t = 71.3). This shows that NEAT captures the context dependence of fixation sequences to an extend that goes beyond word frequency alone.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative Properties",
                "sec_num": "5.2"
            },
            {
                "text": "Parts of Speech Part of speech categories are known to be a predictor of fixation probabilities, with content words being more likely to be fixated than function words (Carpenter and Just, 1983) . In Table 4 , we give the simulated fixation probabilities and the human fixation probabilities estimated from the Dundee corpus for the tags of the Universal PoS tagset (Petrov et al., 2012) , using the PoS annotation of Barrett et al. (2015) . We again compare with the probabilities of a threshold predictor derived from word frequency. 1 NEAT captures the differences between PoS categories well, as evidenced by the high correlation coefficients. The content word categories ADJ, ADV, NOUN, VERB and X consistently show higher probabilities than the function word categories. While the correlation coefficients for word frequency are very similar, the numerical values of the simulated probabilities are closer to the human ones than those derived from word frequency, which tend towards more extreme values. This difference can be seen clearly if we compare the mean squared error, rather than the correlation, with the human fixation probabilities (last row of Table 4 ).",
                "cite_spans": [
                    {
                        "start": 168,
                        "end": 194,
                        "text": "(Carpenter and Just, 1983)",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 366,
                        "end": 387,
                        "text": "(Petrov et al., 2012)",
                        "ref_id": "BIBREF23"
                    },
                    {
                        "start": 418,
                        "end": 439,
                        "text": "Barrett et al. (2015)",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 200,
                        "end": 207,
                        "text": "Table 4",
                        "ref_id": "TABREF7"
                    },
                    {
                        "start": 1164,
                        "end": 1171,
                        "text": "Table 4",
                        "ref_id": "TABREF7"
                    }
                ],
                "eq_spans": [],
                "section": "Qualitative Properties",
                "sec_num": "5.2"
            },
            {
                "text": "Correlations with Known Predictors In the literature, it has been observed that skipping correlates with predictability (surprisal), word frequency, and word length (Rayner, 1998, p. 387) . These correlations are also observed in the human skipping data derived from Dundee, as shown in Table 5 . (Human fixation probabilities were obtained by averaging over the ten readers in Dundee.)",
                "cite_spans": [
                    {
                        "start": 165,
                        "end": 187,
                        "text": "(Rayner, 1998, p. 387)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 287,
                        "end": 294,
                        "text": "Table 5",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Qualitative Properties",
                "sec_num": "5.2"
            },
            {
                "text": "Comparing the known predictors of skipping with NEAT's simulated fixation probabilities, similar correlations as in the human data are observed. We observe that the correlations with surprisal are stronger in NEAT, considering both restricted surprisal and full surprisal as measures of predictability.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative Properties",
                "sec_num": "5.2"
            },
            {
                "text": "We investigated the hypothesis that human reading strategies optimize a tradeoff between precision of language understanding and economy of attention. We made this idea explicit in NEAT, a neural reading architecture with hard attention that can be 1 We omit the tag \".\" for punctuation, as punctuation characters are not treated as separate tokens in Dundee. Table 5 : Correlations between human and NEAT fixation probabilities and known predictors trained end-to-end to optimize this tradeoff. Experiments on the Dundee corpus show that NEAT provides accurate predictions for human skipping behavior. It also predicts reading times, even though it only has access to 60.4% of the words in the corpus in order to estimate surprisal. Finally, we found that known qualitative properties of skipping emerge in our model, even though they were not explicitly included in the architecture, such as context dependence of fixations, differential skipping rates across parts of speech, and correlations with other known predictors of human reading behavior.",
                "cite_spans": [
                    {
                        "start": 249,
                        "end": 250,
                        "text": "1",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 360,
                        "end": 367,
                        "text": "Table 5",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Conclusions",
                "sec_num": "6"
            },
            {
                "text": "The NEAT Reading ModelThe point of departure for our model is the Tradeoff Hypothesis (see Section 1): Reading optimizes a tradeoff between precision of language understand-",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Learning wake-sleep recurrent attention models",
                "authors": [
                    {
                        "first": "Jimmy",
                        "middle": [],
                        "last": "Ba",
                        "suffix": ""
                    },
                    {
                        "first": "Ruslan",
                        "middle": [
                            "R"
                        ],
                        "last": "Salakhutdinov",
                        "suffix": ""
                    },
                    {
                        "first": "Roger",
                        "middle": [
                            "B"
                        ],
                        "last": "Grosse",
                        "suffix": ""
                    },
                    {
                        "first": "Brendan",
                        "middle": [
                            "J"
                        ],
                        "last": "Frey",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Advances in Neural Information Processing Systems",
                "volume": "",
                "issue": "",
                "pages": "2575--2583",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ba, Jimmy, Ruslan R. Salakhutdinov, Roger B. Grosse, and Brendan J. Frey. 2015. Learning wake-sleep recurrent attention models. In Ad- vances in Neural Information Processing Sys- tems. pages 2575-2583.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "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": 2015,
                "venue": "Proceedings of the International Conference on Learning Representations",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Bahdanau, Dzmitry, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In Proceed- ings of the International Conference on Learning Representations.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "The Dundee treebank",
                "authors": [
                    {
                        "first": "Maria",
                        "middle": [],
                        "last": "Barrett",
                        "suffix": ""
                    },
                    {
                        "first": "\u017deljko",
                        "middle": [],
                        "last": "Agi\u0107",
                        "suffix": ""
                    },
                    {
                        "first": "Anders",
                        "middle": [],
                        "last": "S\u00f8gaard",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of the 14th International Workshop on Treebanks and Linguistic Theories",
                "volume": "",
                "issue": "",
                "pages": "242--248",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Barrett, Maria,\u017deljko Agi\u0107, and Anders S\u00f8gaard. 2015. The Dundee treebank. In Proceedings of the 14th International Workshop on Treebanks and Linguistic Theories. pages 242-248.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "A rational model of eye movement control in reading",
                "authors": [
                    {
                        "first": "Klinton",
                        "middle": [],
                        "last": "Bicknell",
                        "suffix": ""
                    },
                    {
                        "first": "Roger",
                        "middle": [],
                        "last": "Levy",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "1168--1178",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Bicknell, Klinton and Roger Levy. 2010. A ratio- nal model of eye movement control in reading. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics. pages 1168-1178.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "What your eyes do while your mind is reading",
                "authors": [
                    {
                        "first": "P",
                        "middle": [
                            "A"
                        ],
                        "last": "Carpenter",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [
                            "A"
                        ],
                        "last": "Just",
                        "suffix": ""
                    }
                ],
                "year": 1983,
                "venue": "Eye Movements in Reading",
                "volume": "",
                "issue": "",
                "pages": "275--307",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Carpenter, P. A. and M. A. Just. 1983. What your eyes do while your mind is reading. In K. Rayner, editor, Eye Movements in Reading, Academic Press, New York, pages 275-307.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Data from eye-tracking corpora as evidence for theories of syntactic processing complexity",
                "authors": [
                    {
                        "first": "Vera",
                        "middle": [],
                        "last": "Demberg",
                        "suffix": ""
                    },
                    {
                        "first": "Frank",
                        "middle": [],
                        "last": "Keller",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Cognition",
                "volume": "109",
                "issue": "2",
                "pages": "193--210",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Demberg, Vera and Frank Keller. 2008. Data from eye-tracking corpora as evidence for theo- ries of syntactic processing complexity. Cognition 109(2):193-210.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "A dynamical model of saccade generation in reading based on spatially distributed lexical processing",
                "authors": [
                    {
                        "first": "Ralf",
                        "middle": [],
                        "last": "Engbert",
                        "suffix": ""
                    },
                    {
                        "first": "Andr\u00e9",
                        "middle": [],
                        "last": "Longtin",
                        "suffix": ""
                    },
                    {
                        "first": "Reinhold",
                        "middle": [],
                        "last": "Kliegl",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "Vision Research",
                "volume": "42",
                "issue": "5",
                "pages": "621--636",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Engbert, Ralf, Andr\u00e9 Longtin, and Reinhold Kliegl. 2002. A dynamical model of saccade generation in reading based on spatially distributed lexical processing. Vision Research 42(5):621-636.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "SWIFT: A dynamical model of saccade generation during reading",
                "authors": [
                    {
                        "first": "Ralf",
                        "middle": [],
                        "last": "Engbert",
                        "suffix": ""
                    },
                    {
                        "first": "Antje",
                        "middle": [],
                        "last": "Nuthmann",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Eike",
                        "suffix": ""
                    },
                    {
                        "first": "Reinhold",
                        "middle": [],
                        "last": "Richter",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Kliegl",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Psychological Review",
                "volume": "112",
                "issue": "4",
                "pages": "777--813",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Engbert, Ralf, Antje Nuthmann, Eike M. Richter, and Reinhold Kliegl. 2005. SWIFT: A dynami- cal model of saccade generation during reading. Psychological Review 112(4):777-813.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Insensitivity of the human sentence-processing system to hierarchical structure",
                "authors": [
                    {
                        "first": "S",
                        "middle": [
                            "L"
                        ],
                        "last": "Frank",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Bod",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Psychological Science",
                "volume": "22",
                "issue": "",
                "pages": "829--834",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Frank, S.L. and R. Bod. 2011. Insensitivity of the human sentence-processing system to hierarchi- cal structure. Psychological Science 22:829-834.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "A probabilistic Earley parser as a psycholinguistic model",
                "authors": [
                    {
                        "first": "John",
                        "middle": [],
                        "last": "Hale",
                        "suffix": ""
                    }
                ],
                "year": 2001,
                "venue": "Proceedings of Conference of the North American Chapter of the Association for Computational Linguistics",
                "volume": "2",
                "issue": "",
                "pages": "159--166",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hale, John. 2001. A probabilistic Earley parser as a psycholinguistic model. In Proceedings of Con- ference of the North American Chapter of the Association for Computational Linguistics. vol- ume 2, pages 159-166.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Predicting word fixations in text with a CRF model for capturing general reading strategies among readers",
                "authors": [
                    {
                        "first": "Tadayoshi",
                        "middle": [],
                        "last": "Hara",
                        "suffix": ""
                    },
                    {
                        "first": "Daichi",
                        "middle": [],
                        "last": "Mochihashi Yoshinobu",
                        "suffix": ""
                    },
                    {
                        "first": "Akiko",
                        "middle": [],
                        "last": "Kano",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Aizawa",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "Proceedings of the 1st Workshop on Eye-tracking and Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "55--70",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hara, Tadayoshi, Daichi Mochihashi Yoshinobu Kano, and Akiko Aizawa. 2012. Predicting word fixations in text with a CRF model for capturing general reading strategies among readers. In Pro- ceedings of the 1st Workshop on Eye-tracking and Natural Language Processing. pages 55-70.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Human gaze control in realworld scene perception",
                "authors": [
                    {
                        "first": "John",
                        "middle": [],
                        "last": "Henderson",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Trends in Cognitive Sciences",
                "volume": "7",
                "issue": "",
                "pages": "498--504",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Henderson, John. 2003. Human gaze control in real- world scene perception. Trends in Cognitive Sci- ences 7:498-504.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Teaching machines to read and comprehend",
                "authors": [
                    {
                        "first": "Karl",
                        "middle": [],
                        "last": "Hermann",
                        "suffix": ""
                    },
                    {
                        "first": "Tom\u00e1\u0161",
                        "middle": [],
                        "last": "Moritz",
                        "suffix": ""
                    },
                    {
                        "first": "Edward",
                        "middle": [],
                        "last": "Ko\u010disk\u1ef3",
                        "suffix": ""
                    },
                    {
                        "first": "Lasse",
                        "middle": [],
                        "last": "Grefenstette",
                        "suffix": ""
                    },
                    {
                        "first": "Will",
                        "middle": [],
                        "last": "Espeholt",
                        "suffix": ""
                    },
                    {
                        "first": "Mustafa",
                        "middle": [],
                        "last": "Kay",
                        "suffix": ""
                    },
                    {
                        "first": "Phil",
                        "middle": [],
                        "last": "Suleyman",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Blunsom",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hermann, Karl Moritz, Tom\u00e1\u0161 Ko\u010disk\u1ef3, Ed- ward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. ArXiv:1506.03340.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Long short-term memory",
                "authors": [
                    {
                        "first": "Sepp",
                        "middle": [],
                        "last": "Hochreiter",
                        "suffix": ""
                    },
                    {
                        "first": "J\u00fcrgen",
                        "middle": [],
                        "last": "Schmidhuber",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "Neural Computation",
                "volume": "9",
                "issue": "8",
                "pages": "1735--1780",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hochreiter, Sepp and J\u00fcrgen Schmidhuber. 1997. Long short-term memory. Neural Computation 9(8):1735-1780.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Parafoveal-onfoveal effects in normal reading",
                "authors": [
                    {
                        "first": "Alan",
                        "middle": [],
                        "last": "Kennedy",
                        "suffix": ""
                    },
                    {
                        "first": "Jo\u00ebl",
                        "middle": [],
                        "last": "Pynte",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Vision Research",
                "volume": "45",
                "issue": "2",
                "pages": "153--168",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kennedy, Alan and Jo\u00ebl Pynte. 2005. Parafoveal-on- foveal effects in normal reading. Vision Research 45(2):153-168.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Expectation-based syntactic comprehension",
                "authors": [
                    {
                        "first": "Roger",
                        "middle": [],
                        "last": "Levy",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Cognition",
                "volume": "106",
                "issue": "3",
                "pages": "1126--1177",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Levy, Roger. 2008. Expectation-based syntactic comprehension. Cognition 106(3):1126-1177.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "With blinkers on: Robust prediction of eye movements across readers",
                "authors": [
                    {
                        "first": "Franz",
                        "middle": [],
                        "last": "Matthies",
                        "suffix": ""
                    },
                    {
                        "first": "Anders",
                        "middle": [],
                        "last": "S\u00f8gaard",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of the Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "803--807",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Matthies, Franz and Anders S\u00f8gaard. 2013. With blinkers on: Robust prediction of eye movements across readers. In Proceedings of the Conference on Empirical Methods in Natural Language Pro- cessing. pages 803-807.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Eye movements reveal the on-line computation of lexical probabilities during reading",
                "authors": [
                    {
                        "first": "Scott",
                        "middle": [
                            "A"
                        ],
                        "last": "Mcdonald",
                        "suffix": ""
                    },
                    {
                        "first": "Richard",
                        "middle": [
                            "C"
                        ],
                        "last": "Shillcock",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Psychological Science",
                "volume": "14",
                "issue": "6",
                "pages": "648--652",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "McDonald, Scott A. and Richard C. Shillcock. 2003a. Eye movements reveal the on-line compu- tation of lexical probabilities during reading. Psy- chological Science 14(6):648-652.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Low-level predictive inference in reading: the influence of transitional probabilities on eye movements",
                "authors": [
                    {
                        "first": "Scott",
                        "middle": [
                            "A"
                        ],
                        "last": "Mcdonald",
                        "suffix": ""
                    },
                    {
                        "first": "Richard",
                        "middle": [
                            "C"
                        ],
                        "last": "Shillcock",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Vision Research",
                "volume": "43",
                "issue": "16",
                "pages": "1735--1751",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "McDonald, Scott A. and Richard C. Shillcock. 2003b. Low-level predictive inference in reading: the influence of transitional probabilities on eye movements. Vision Research 43(16):1735-1751.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "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\u00fd",
                        "suffix": ""
                    },
                    {
                        "first": "Sanjeev",
                        "middle": [],
                        "last": "Khudanpur",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of Interspeech",
                "volume": "",
                "issue": "",
                "pages": "1045--1048",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mikolov, Tom\u00e1\u0161, Martin Karafi\u00e1t, Luk\u00e1\u0161 Burget, Ja\u0148 Cernock\u00fd, and Sanjeev Khudanpur. 2010. Recur- rent neural network based language model. In Proceedings of Interspeech. pages 1045-1048.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Recurrent models of visual attention",
                "authors": [
                    {
                        "first": "Volodymyr",
                        "middle": [],
                        "last": "Mnih",
                        "suffix": ""
                    },
                    {
                        "first": "Nicolas",
                        "middle": [],
                        "last": "Heess",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Advances in Neural Information Processing Systems",
                "volume": "",
                "issue": "",
                "pages": "2204--2212",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mnih, Volodymyr, Nicolas Heess, Alex Graves, and others. 2014. Recurrent models of visual atten- tion. In Advances in Neural Information Process- ing Systems. pages 2204-2212.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Learning where to look: Modeling eye movements in reading",
                "authors": [
                    {
                        "first": "Mattias",
                        "middle": [],
                        "last": "Nilsson",
                        "suffix": ""
                    },
                    {
                        "first": "Joakim",
                        "middle": [],
                        "last": "Nivre",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Proceedings of the 13th Conference on Computational Natural Language Learning",
                "volume": "",
                "issue": "",
                "pages": "93--101",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Nilsson, Mattias and Joakim Nivre. 2009. Learn- ing where to look: Modeling eye movements in reading. In Proceedings of the 13th Conference on Computational Natural Language Learning. pages 93-101.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "Towards a data-driven model of eye movement control in reading",
                "authors": [
                    {
                        "first": "Mattias",
                        "middle": [],
                        "last": "Nilsson",
                        "suffix": ""
                    },
                    {
                        "first": "Joakim",
                        "middle": [],
                        "last": "Nivre",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "63--71",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Nilsson, Mattias and Joakim Nivre. 2010. Towards a data-driven model of eye movement control in reading. In Proceedings of the Workshop on Cog- nitive Modeling and Computational Linguistics. pages 63-71.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "A universal part-of-speech tagset",
                "authors": [
                    {
                        "first": "Slav",
                        "middle": [],
                        "last": "Petrov",
                        "suffix": ""
                    },
                    {
                        "first": "Dipanjan",
                        "middle": [],
                        "last": "Das",
                        "suffix": ""
                    },
                    {
                        "first": "Ryan",
                        "middle": [
                            "T"
                        ],
                        "last": "Mcdonald",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "Proceedings of the 8th International Conference on Language Resources and Evaluation",
                "volume": "",
                "issue": "",
                "pages": "2089--2096",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Petrov, Slav, Dipanjan Das, and Ryan T. McDon- ald. 2012. A universal part-of-speech tagset. In Proceedings of the 8th International Conference on Language Resources and Evaluation. pages 2089-2096.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "Eye movements in reading and information processing: 20 years of research",
                "authors": [
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Rayner",
                        "suffix": ""
                    }
                ],
                "year": 1998,
                "venue": "Psychological Bulletin",
                "volume": "124",
                "issue": "3",
                "pages": "372--422",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Rayner, K. 1998. Eye movements in reading and in- formation processing: 20 years of research. Psy- chological Bulletin 124(3):372-422.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "title": "Eye movements in reading: Models and data",
                "authors": [
                    {
                        "first": "Keith",
                        "middle": [],
                        "last": "Rayner",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Journal of Eye Movement Research",
                "volume": "2",
                "issue": "5",
                "pages": "1--10",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Rayner, Keith. 2009. Eye movements in reading: Models and data. Journal of Eye Movement Re- search 2(5):1-10.",
                "links": null
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "Models of the reading process",
                "authors": [
                    {
                        "first": "Keith",
                        "middle": [],
                        "last": "Rayner",
                        "suffix": ""
                    },
                    {
                        "first": "Erik",
                        "middle": [
                            "D"
                        ],
                        "last": "Reichle",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Cognitive Science",
                "volume": "1",
                "issue": "6",
                "pages": "787--799",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Rayner, Keith and Erik D. Reichle. 2010. Models of the reading process. Wiley Interdisciplinary Re- views: Cognitive Science 1(6):787-799.",
                "links": null
            },
            "BIBREF27": {
                "ref_id": "b27",
                "title": "Toward a model of eye movement control in reading",
                "authors": [
                    {
                        "first": "E",
                        "middle": [
                            "D"
                        ],
                        "last": "Reichle",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Pollatsek",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [
                            "L"
                        ],
                        "last": "Fisher",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Rayner",
                        "suffix": ""
                    }
                ],
                "year": 1998,
                "venue": "Psychological Review",
                "volume": "105",
                "issue": "1",
                "pages": "125--157",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Reichle, E. D., A. Pollatsek, D. L. Fisher, and K. Rayner. 1998. Toward a model of eye move- ment control in reading. Psychological Review 105(1):125-157.",
                "links": null
            },
            "BIBREF28": {
                "ref_id": "b28",
                "title": "Using EZ Reader to model the effects of higher level language processing on eye movements during reading",
                "authors": [
                    {
                        "first": "E",
                        "middle": [
                            "D"
                        ],
                        "last": "Reichle",
                        "suffix": ""
                    },
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Warren",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Mcconnell",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Psychonomic Bulletin & Review",
                "volume": "16",
                "issue": "1",
                "pages": "1--21",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Reichle, E. D., T. Warren, and K. McConnell. 2009. Using EZ Reader to model the effects of higher level language processing on eye movements dur- ing reading. Psychonomic Bulletin & Review 16(1):1-21.",
                "links": null
            },
            "BIBREF29": {
                "ref_id": "b29",
                "title": "The EZ Reader model of eyemovement control in reading: Comparisons to other models",
                "authors": [
                    {
                        "first": "Erik",
                        "middle": [
                            "D"
                        ],
                        "last": "Reichle",
                        "suffix": ""
                    },
                    {
                        "first": "Keith",
                        "middle": [],
                        "last": "Rayner",
                        "suffix": ""
                    },
                    {
                        "first": "Alexander",
                        "middle": [],
                        "last": "Pollatsek",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Behavioral and Brain Sciences",
                "volume": "26",
                "issue": "04",
                "pages": "445--476",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Reichle, Erik D., Keith Rayner, and Alexander Pol- latsek. 2003. The EZ Reader model of eye- movement control in reading: Comparisons to other models. Behavioral and Brain Sciences 26(04):445-476.",
                "links": null
            },
            "BIBREF30": {
                "ref_id": "b30",
                "title": "The effect of word predictability on reading time is logarithmic",
                "authors": [
                    {
                        "first": "Nathaniel",
                        "middle": [
                            "J"
                        ],
                        "last": "Smith",
                        "suffix": ""
                    },
                    {
                        "first": "Roger",
                        "middle": [],
                        "last": "Levy",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Cognition",
                "volume": "128",
                "issue": "3",
                "pages": "302--319",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Smith, Nathaniel J. and Roger Levy. 2013. The ef- fect of word predictability on reading time is log- arithmic. Cognition 128(3):302-319.",
                "links": null
            },
            "BIBREF31": {
                "ref_id": "b31",
                "title": "Sequence to sequence learning with neural networks",
                "authors": [
                    {
                        "first": "Ilya",
                        "middle": [],
                        "last": "Sutskever",
                        "suffix": ""
                    },
                    {
                        "first": "Oriol",
                        "middle": [],
                        "last": "Vinyals",
                        "suffix": ""
                    },
                    {
                        "first": "Quoc Vv",
                        "middle": [],
                        "last": "Le",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Advances in Neural Information Processing Systems",
                "volume": "",
                "issue": "",
                "pages": "3104--3112",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sutskever, Ilya, Oriol Vinyals, and Quoc VV Le. 2014. Sequence to sequence learning with neu- ral networks. In Advances in Neural Information Processing Systems. pages 3104-3112.",
                "links": null
            },
            "BIBREF32": {
                "ref_id": "b32",
                "title": "Syntactic parsing",
                "authors": [
                    {
                        "first": "Roger",
                        "middle": [],
                        "last": "Van Gompel",
                        "suffix": ""
                    },
                    {
                        "first": "Martin",
                        "middle": [
                            "J"
                        ],
                        "last": "Pg",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Pickering",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "The Oxford Handbook of Psycholinguistics",
                "volume": "",
                "issue": "",
                "pages": "289--307",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Van Gompel, Roger PG and Martin J. Pickering. 2007. Syntactic parsing. In The Oxford Handbook of Psycholinguistics, Oxford University Press, pages 289-307.",
                "links": null
            },
            "BIBREF33": {
                "ref_id": "b33",
                "title": "Simple statistical gradient-following algorithms for connectionist reinforcement learning",
                "authors": [
                    {
                        "first": "Ronald",
                        "middle": [
                            "J"
                        ],
                        "last": "Williams",
                        "suffix": ""
                    }
                ],
                "year": 1992,
                "venue": "Machine Learning",
                "volume": "8",
                "issue": "3-4",
                "pages": "229--256",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Williams, Ronald J. 1992. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning 8(3- 4):229-256.",
                "links": null
            },
            "BIBREF34": {
                "ref_id": "b34",
                "title": "Show, attend and tell: Neural image caption generation with visual attention",
                "authors": [
                    {
                        "first": "Kelvin",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    },
                    {
                        "first": "Jimmy",
                        "middle": [],
                        "last": "Ba",
                        "suffix": ""
                    },
                    {
                        "first": "Ryan",
                        "middle": [],
                        "last": "Kiros",
                        "suffix": ""
                    },
                    {
                        "first": "Aaron",
                        "middle": [],
                        "last": "Courville",
                        "suffix": ""
                    },
                    {
                        "first": "Ruslan",
                        "middle": [],
                        "last": "Salakhutdinov",
                        "suffix": ""
                    },
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Zemel",
                        "suffix": ""
                    },
                    {
                        "first": "Yoshua",
                        "middle": [],
                        "last": "Bengio",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xu, Kelvin, Jimmy Ba, Ryan Kiros, Aaron Courville, Ruslan Salakhutdinov, Richard Zemel, and Yoshua Bengio. 2015. Show, attend and tell: Neural image caption generation with visual at- tention. ArXiv:1502.03044.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "num": null,
                "text": "Top: Heatmap showing human fixation probabilities, as estimated from the ten readers in the Dundee corpus. In cases of track loss, we replaced the missing value with the corresponding reader's overall fixation rate. Bottom: Heatmap showing fixation probabilities simulated by NEAT. Color gradient ranges from blue (low probability) to red (high probability); words without color are at the beginning or end of a sequence, or out of vocabulary.",
                "uris": null,
                "type_str": "figure"
            },
            "TABREF1": {
                "num": null,
                "html": null,
                "type_str": "table",
                "text": "Evaluation of fixation sequence predictions against hu-",
                "content": "<table><tr><td>man data. For the human baseline, we predicted the n-th reader's</td></tr><tr><td>fixations by taking the fixations of the n + 1-th reader (with</td></tr><tr><td>missing values replaced by reader average), averaging the re-</td></tr><tr><td>sulting scores over the ten readers.</td></tr></table>"
            },
            "TABREF3": {
                "num": null,
                "html": null,
                "type_str": "table",
                "text": "Performance on language modeling and reconstruction as measured by perplexity. Random attention is an upper bound on perplexity, while full attention is a lower bound. For the human baseline, we give two figures, which differ in the treatment of missing data. The first figure is obtained when replacing missing values with a random variable \u03c9 \u223c Binom(n, 0.61); the second results from replacing missing values with 1.",
                "content": "<table/>"
            },
            "TABREF5": {
                "num": null,
                "html": null,
                "type_str": "table",
                "text": "Linear mixed effects models for first pass duration.",
                "content": "<table><tr><td>The first part of the table shows the coefficients, standard er-</td></tr><tr><td>rors, and t values for the predictors in the baseline model. The</td></tr><tr><td>second part of the table gives the corresponding values for ran-</td></tr><tr><td>dom surprisal, restricted surprisal computed by NEAT, and full</td></tr><tr><td>surprisal, residualized against the baseline predictors, in three</td></tr><tr><td>models obtained by adding these predictors.</td></tr></table>"
            },
            "TABREF7": {
                "num": null,
                "html": null,
                "type_str": "table",
                "text": "Actual and simulated fixation probabilities (in %) byPoS tag, with the ranks given in brackets, and correlations and mean squared error relative to human data.",
                "content": "<table/>"
            }
        }
    }
}