Unsupervised Recurrent Neural Network Grammars

  • 2019-08-05 01:21:15
  • Yoon Kim, Alexander M. Rush, Lei Yu, Adhiguna Kuncoro, Chris Dyer, Gábor Melis
  • 0

Abstract

Recurrent neural network grammars (RNNG) are generative models of languagewhich jointly model syntax and surface structure by incrementally generating asyntax tree and sentence in a top-down, left-to-right order. Supervised RNNGsachieve strong language modeling and parsing performance, but require anannotated corpus of parse trees. In this work, we experiment with unsupervisedlearning of RNNGs. Since directly marginalizing over the space of latent treesis intractable, we instead apply amortized variational inference. To maximizethe evidence lower bound, we develop an inference network parameterized as aneural CRF constituency parser. On language modeling, unsupervised RNNGsperform as well their supervised counterparts on benchmarks in English andChinese. On constituency grammar induction, they are competitive with recentneural language models that induce tree structures from words through attentionmechanisms.

 

Quick Read (beta)

Unsupervised Recurrent Neural Network Grammars

Yoon Kim   Alexander M. Rush   Lei Yu
Adhiguna Kuncoro,   Chris Dyer   Gábor Melis

Harvard University   University of Oxford   DeepMind
{yoonkim,srush}@seas.harvard.edu   {leiyu,akuncoro,cdyer,melisgl}@google.com
Abstract

Recurrent neural network grammars (RNNG) are generative models of language which jointly model syntax and surface structure by incrementally generating a syntax tree and sentence in a top-down, left-to-right order. Supervised RNNGs achieve strong language modeling and parsing performance, but require an annotated corpus of parse trees. In this work, we experiment with unsupervised learning of RNNGs. Since directly marginalizing over the space of latent trees is intractable, we instead apply amortized variational inference. To maximize the evidence lower bound, we develop an inference network parameterized as a neural CRF constituency parser. On language modeling, unsupervised RNNGs perform as well their supervised counterparts on benchmarks in English and Chinese. On constituency grammar induction, they are competitive with recent neural language models that induce tree structures from words through attention mechanisms.

\crefformat

section§#2#1#3 \crefformatsubsection§#2#1#3 \crefformatsubsubsection§#2#1#3 \usetikzlibrarypatterns \algrenewcommand\algorithmicindent1em \algrenewcommand\[email protected]

Unsupervised Recurrent Neural Network Grammars


Yoon Kim   Alexander M. Rush   Lei Yu Adhiguna Kuncoro,   Chris Dyer   Gábor Melis Harvard University   University of Oxford   DeepMind {yoonkim,srush}@seas.harvard.edu   {leiyu,akuncoro,cdyer,melisgl}@google.com

1 Introduction

Work done while the first author was an intern at DeepMind. Code available at https://github.com/harvardnlp/urnng

Recurrent neural network grammars (RNNGs) (Dyer et al., 2016) model sentences by first generating a nested, hierarchical syntactic structure which is used to construct a context representation to be conditioned upon for upcoming words. Supervised RNNGs have been shown to outperform standard sequential language models, achieve excellent results on parsing (Dyer et al., 2016; Kuncoro et al., 2017), better encode syntactic properties of language (Kuncoro et al., 2018), and correlate with electrophysiological responses in the human brain (Hale et al., 2018). However, these all require annotated syntactic trees for training. In this work, we explore unsupervised learning of recurrent neural network grammars for language modeling and grammar induction.

The standard setup for unsupervised structure learning is to define a generative model pθ(𝐱,𝐳) over observed data 𝐱 (e.g. sentence) and unobserved structure 𝐳 (e.g. parse tree, part-of-speech sequence), and maximize the log marginal likelihood logpθ(𝐱)=log𝐳pθ(𝐱,𝐳). Successful approaches to unsupervised parsing have made strong conditional independence assumptions (e.g. context-freeness) and employed auxiliary objectives (Klein and Manning, 2002) or priors (Johnson et al., 2007). These strategies imbue the learning process with inductive biases that guide the model to discover meaningful structures while allowing tractable algorithms for marginalization; however, they come at the expense of language modeling performance, particularly compared to sequential neural models that make no independence assumptions.

Like RNN language models, RNNGs make no independence assumptions. Instead they encode structural bias through operations that compose linguistic constituents. The lack of independence assumptions contributes to the strong language modeling performance of RNNGs, but make unsupervised learning challenging. First, marginalization is intractable. Second, the biases imposed by the RNNG are relatively weak compared to those imposed by models like PCFGs. There is little pressure for non-trivial tree structure to emerge during unsupervised RNNG (URNNG) learning.

In this work, we explore a technique for handling intractable marginalization while also injecting inductive bias. Specifically we employ amortized variational inference (Kingma and Welling, 2014; Rezende et al., 2014; Mnih and Gregor, 2014) with a structured inference network. Variational inference lets us tractably optimize a lower bound on the log marginal likelihood, while employing a structured inference network encourages non-trivial structure. In particular, a conditional random field (CRF) constituency parser (Finkel et al., 2008; Durrett and Klein, 2015), which makes significant independence assumptions, acts as a guide on the generative model to learn meaningful trees through regularizing the posterior (Ganchev et al., 2010).

We experiment with URNNGs on English and Chinese and observe that they perform well as language models compared to their supervised counterparts and standard neural LMs. In terms of grammar induction, they are competitive with recently-proposed neural architectures that discover tree-like structures through gated attention (Shen et al., 2018). Our results, along with other recent work on joint language modeling/structure learning with deep networks (Shen et al., 2018, 2019; Wiseman et al., 2018; Kawakami et al., 2018), suggest that it is possible to learn generative models of language that model the underlying data well (i.e. assign high likelihood to held-out data) and at the same time induce meaningful linguistic structure.

2 Unsupervised Recurrent Neural Network Grammars

We use 𝐱=[x1,,xT] to denote a sentence of length T, and 𝐳𝒵T to denote an unlabeled binary parse tree over a sequence of length T, represented as a binary vector of length 2T-1. Here 0 and 1 correspond to shift and reduce actions, explained below.11 1 The cardinality of 𝒵T{0,1}2T-1 is given by the (T-1)-th Catalan number, |𝒵T|=(2T-2)!T!(T-1)!. Figure 1 presents an overview of our approach.

2.1 Generative Model

An RNNG defines a joint probability distribution pθ(𝐱,𝐳) over sentences 𝐱 and parse trees 𝐳. We consider a simplified version of the original RNNG (Dyer et al., 2016) by ignoring constituent labels and only considering binary trees. The RNNG utilizes an RNN to parameterize a stack data structure (Dyer et al., 2015) of partially-completed constituents to incrementally build the parse tree while generating terminals. Using the current stack representation, the model samples an action (shift or reduce): shift generates a terminal symbol, i.e. word, and shifts it onto the stack,22 2 A better name for shift would be generate (as in Dyer et al. (2016)), but we use shift to emphasize similarity with the shift-reduce parsing. reduce pops the last two elements off the stack, composes them, and shifts the composed representation onto the stack.

Figure 1: Overview of our approach. The inference network qϕ(𝐳|𝐱) (left) is a CRF parser which produces a distribution over binary trees (shown in dotted box). 𝐁ij are random variables for existence of a constituent spanning i-th and j-th words, whose potentials are the output from a bidirectional LSTM (the global factor ensures that the distribution is only over valid binary trees). The generative model pθ(𝐱,𝐳) (right) is an RNNG which consists of a stack LSTM (from which actions/words are predicted) and a tree LSTM (to obtain constituent representations upon reduce). Training involves sampling a binary tree from qϕ(𝐳|𝐱), converting it to a sequence of shift/reduce actions (𝐳= [shift, shift, shift, reduce, reduce, shift, reduce] in the above example), and optimizing the log joint likelihood logpθ(𝐱,𝐳).

Formally, let S=[(𝟎,𝟎)] be the initial stack. Each item of the stack will be a pair, where the first element is the hidden state of the stack LSTM, and the second element is an input vector, described below. We use top(S) to refer to the top pair in the stack. The push and pop operations are defined imperatively in the usual way. At each time step, the next action zt (shift or reduce) is sampled from a Bernoulli distribution parameterized in terms of the current stack representation. Letting (𝐡prev,𝐠prev)=top(S), we have

ztBernoulli(pt),pt=σ(𝐰𝐡prev+b).

Subsequent generation depend on zt:

  • If zt=0 (shift), the model first generates a terminal symbol via sampling from a categorical distribution whose parameters come from an affine transformation and a softmax,

    xsoftmax(𝐖𝐡prev+𝐛).

    Then the generated terminal is shifted onto the stack using a stack LSTM,

    𝐡next=LSTM(𝐞x,𝐡prev),
    push(S,(𝐡next,𝐞x)),

    where 𝐞x is the word embedding for x.

  • If zt=1 (reduce), we pop the last two elements off the stack,

    (𝐡r,𝐠r)=pop(S),(𝐡l,𝐠l)=pop(S),

    and obtain a new representation that combines the left/right constituent representations using a tree LSTM (Tai et al., 2015; Zhu et al., 2015),

    𝐠new=TreeLSTM(𝐠l,𝐠r).

    Note that we use 𝐠l and 𝐠r to obtain the new representation instead of 𝐡l and 𝐡r.33 3 The update equations for the tree LSTM (and the stack LSTM) also involve cell states in addition to the hidden states. To reduce notational clutter we do not explicitly show the cell states and instead subsume them into 𝐠. If one (or both) of the inputs to the tree LSTM is a word embedding, the associated cell state is taken to be zero. See Tai et al. (2015) for the exact parameterization. We then update the stack using 𝐠new,

    (𝐡prev,𝐠prev)=top(S),
    𝐡new=LSTM(𝐠new,𝐡prev),
    push(S,(𝐡new,𝐠new)).

The generation process continues until an end-of-sentence symbol is generated. The parameters θ of the generative model are 𝐰,b,𝐖,𝐛, and the parameters of the stack/tree LSTMs. For a sentence 𝐱=[x1,,xT] of length T, the binary parse tree is given by the binary vector 𝐳=[z1,,z2T-1].44 4 As it stands, the support of 𝐳 is {0,1}2T-1, all binary vectors of length 2T-1. To restrict our distribution to 𝒵T (binary vectors which describe valid trees), we constrain zt to be valid at each time step, which amounts to deterministically choosing zt=0 (shift) if there are fewer than two elements (not counting the initial zero tuple) on the stack. The joint log likelihood decomposes as a sum of terminal/action log likelihoods,

logpθ( 𝐱,𝐳)=t=1Tlogpθ(xt|𝐱<t,𝐳<n(t))logpθ(𝐱|𝐳)
+j=12T-1logpθ(zj|𝐱<m(j),𝐳<j)logpθ(𝐳|𝐱<𝐳), (1)

where 𝐳<n(t) refers to all actions before generating the t-th word, and similarly 𝕩<m(j) refers to all words generated before taking the j-th action. For brevity, from here on we will use logpθ(𝐱|𝐳) to refer to the first term (terminal log likelihood) and logpθ(𝐳|𝐱<𝐳) to refer to the second term (action log likelihood) in the above decomposition.55 5 The action log likelihood is the sum of log conditional priors, which is obviously different from the unconditional log prior logpθ(𝐳)=log𝐱pθ(𝐱,𝐳).

In the supervised case where ground-truth 𝐳 is available, we can straightforwardly perform gradient-based optimization to maximize the joint log likelihood logpθ(𝐱,𝐳). In the unsupervised case, the standard approach is to maximize the log marginal likelihood,

logpθ(𝐱)=log𝐳𝒵Tpθ(𝐱,𝐳).

However this summation is intractable because zt fully depends on all previous actions [z1,,zt-1]. Even if this summation were tractable, it is not clear that meaningful latent structures would emerge given the lack of explicit independence assumptions in the RNNG (e.g. it is clearly not context-free). We handle these issues with amortized variational inference.

2.2 Amortized Variational Inference

Amortized variational inference (Kingma and Welling, 2014) defines a trainable inference network ϕ that parameterizes qϕ(𝐳|𝐱), a variational posterior distribution, in this case over parse trees 𝐳 given the sentence 𝐱. This distribution is used to form an evidence lower bound (ELBO) on the log marginal likelihood,

ELBO(θ,ϕ;𝐱)=𝔼qϕ(𝐳|𝐱)[logpθ(𝐱,𝐳)qϕ(𝐳|𝐱)].

We maximize the ELBO with respect to both model parameters θ and inference network parameters ϕ. The ELBO is still intractable to calculate exactly, but this formulation will allow us to obtain unbiased gradient estimators based on Monte Carlo sampling.

Observe that rearranging the ELBO gives the following optimization problem,

maxθ,ϕlogpθ(𝐱)-KL[qϕ(𝐳|𝐱)pθ(𝐳|𝐱)].

Thus, ϕ is trained to match the variational posterior qϕ(𝐳|𝐱) to the true posterior pθ(𝐳|𝐱), but θ is also trained to match the true posterior to the variational posterior. Indeed, there is some evidence to suggest that generative models trained with amortized variational inference (i.e. variational autoencoders) learn posterior distributions that are close to the variational family (Cremer et al., 2018).

We can use this to our advantage with an inference network that injects inductive bias. We propose to do this by using a context-free model for the inference network, in particular, a neural CRF parser (Durrett and Klein, 2015). This choice can seen as a form of posterior regularization that limits posterior flexibility of the overly powerful RNNG generative model.66 6 While it has a similar goal, this formulation differs the from posterior regularization as formulated by Ganchev et al. (2010), which constrains the distributional family via linear constraints on posterior expectations. In our case, the conditional independence assumptions in the CRF lead to a curved exponential family where the vector of natural parameters has fewer dimensions than the vector of sufficient statistics of the full exponential family. This curved exponential family is a subset of the marginal polytope of the full exponential family, but it is an intersection of both linear and nonlinear manifolds, and therefore cannot be characterized through linear constraints over posterior expectations.,77 7 In preliminary experiments, we also attempted to learn latent trees with a transition-based parser (which does not make explicit independence assumptions) that looks at the entire sentence. However we found that under this setup, the inference network degenerated into a local minimum whereby it always generated left-branching trees despite various optimization strategies. Williams et al. (2018) observe a similar phenomenon in the context of learning latent trees for classification tasks. However Li et al. (2019) find that it is possible use a transition-based parser as the inference network for dependency grammar induction, if the inference network is constrained via posterior regularization (Ganchev et al., 2010) based on universal syntactic rules (Naseem et al., 2010).

The parameterization of span scores is similar to recent works (Wang and Chang, 2016; Stern et al., 2017; Kitaev and Klein, 2018): we add position embeddings to word embeddings and run a bidirectional LSTM over the input representations to obtain the forward [𝐡1,,𝐡T] and backward [𝐡1,,𝐡T] hidden states. The score sij for a constituent spanning xi to xj is given by,

sij=MLP([𝐡j+1-𝐡i;𝐡i-1-𝐡j]).

Letting 𝐁 be the binary matrix representation of a tree (𝐁ij=1 means there is a constituent spanning xi and xj), the CRF parser defines a distribution over binary trees via the Gibbs distribution,

qϕ(𝐁|𝐱)=1ZT(𝐱)exp(ij𝐁ijsij),

where ZT(𝐱) is the partition function,

ZT(𝐱)=𝐁Texp(ij𝐁ijsij),

and ϕ denotes the parameters of the inference network (i.e. the bidirectional LSTM and the MLP). Calculating ZT(𝐱) requires a summation over an exponentially-sized set T{0,1}T×T, the set of all binary trees over a length T sequence. However we can perform the summation in O(T3) using the inside algorithm (Baker, 1979), shown in Algorithm 2.2. This computation is itself differentiable and amenable to gradient-based optimization. Finally, letting f:T𝒵T be the bijection between the binary tree matrix representation and a sequence of shift/reduce actions, the inference network defines a distribution over 𝒵T via qϕ(𝐳|𝐱)qϕ(f-1(𝐳)|𝐱). {algorithm}[t] Inside algorithm for calculating ZT(𝐱) {algorithmic}[1] \ProcedureInsides \Commentscores sij for ij \Fori:=1 to T \Commentlength-1 spans \Stateβ[i,i]=exp(sii) \EndFor\For:=1 to T-1 \Commentspan length \Fori:=1 to T- \Commentspan start \Statej=i+ \Commentspan end \Stateβ[i,j]=k=ij-1exp(sij)β[i,k]β[k+1,j] \EndFor\EndFor\Statereturn β[1,T]\Commentreturn partition function ZT(𝐱) \EndProcedure

2.3 Optimization

For optimization, we use the following variant of the ELBO,

𝔼qϕ(𝐳|𝐱)[logpθ(𝐱,𝐳)]+[qϕ(𝐳|𝐱)],

where [qϕ(𝐳|𝐱)]=𝔼qϕ(𝐳|𝐱)[-logqϕ(𝐳|𝐱)] is the entropy of the variational posterior. A Monte Carlo estimate for the gradient with respect to θ is

θELBO(θ,ϕ;𝐱)1Kk=1Kθlogpθ(𝐱,𝐳(k)),

with samples 𝐳(1),,𝐳(K) from qϕ(𝐳|𝐱). Sampling uses the intermediate values calculated during the inside algorithm to sample split points recursively (Goodman, 1998; Finkel et al., 2006), as shown in Algorithm 2.3. The gradient with respect to ϕ involves two parts. The entropy term [qϕ(𝐳|𝐱)] can be calculated exactly in O(T3), again using the intermediate values from the inside algorithm (see Algorithm 3.1).88 8 We adapt the algorithm for calculating tree entropy in PCFGs from Hwa (2000) to the CRF case. Since each step of this dynamic program is differentiable, we can obtain the gradient ϕ[qϕ(𝐳|𝐱)] using automatic differentation.99 9 ϕ[qϕ(𝐳|𝐱)] can also be computed using the inside-outside algorithm and a second-order expectation semiring (Li and Eisner, 2009), which has the same asymptotic runtime complexity but generally better constants. An estimator for the gradient with respect to 𝔼qϕ(𝐳|𝐱)[logpθ(𝐱,𝐳)] is obtained via the score function gradient estimator (Glynn, 1987; Williams, 1992),

ϕ 𝔼qϕ(𝐳|𝐱)[logpθ(𝐱,𝐳)]
=𝔼qϕ(𝐳|𝐱)[logpθ(𝐱,𝐳)ϕlogqϕ(𝐳|𝐱)]
1Kk=1Klogpθ(𝐱,𝐳(k))ϕlogqϕ(𝐳(k)|𝐱).

The above estimator is unbiased but typically suffers from high variance. To reduce variance, we use a control variate derived from an average of the other samples’ joint likelihoods (Mnih and Rezende, 2016), yielding the following estimator,

1Kk=1K(logpθ(𝐱,𝐳(k))-r(k))ϕlogqϕ(𝐳(k)|𝐱),

where r(k)=1K-1jklogpθ(𝐱,𝐳(j)). This control variate worked better than alternatives such as estimates of baselines from an auxiliary network (Mnih and Gregor, 2014; Deng et al., 2018) or a language model (Yin et al., 2018).

{algorithm}

[t] Top-down sampling a tree from qϕ(𝐳|𝐱) {algorithmic}[1] \ProcedureSampleβ \Commentβ from running Inside(s) \State𝐁=𝟎 \Commentbinary matrix representation of tree \StateQ=[(1,T)] \Commentqueue of constituents \WhileQ is not empty \State(i,j)=pop(Q) \Stateτ=k=ij-1β[i,k]β[k+1,j] \Fork:=i to j-1 \Commentget distribution over splits \Statewk=(β[i,k]β[k+1,j])/τ \EndFor\StatekCat([wi,,wj-1]) \Commentsample a split point \State𝐁i,k=1,𝐁k+1,j=1 \Commentupdate 𝐁 \Ifk>i \Commentif left child has width > 1 \Statepush(Q,(i,k)) \Commentadd to queue \EndIf\Ifk+1<j \Commentif right child has width > 1 \Statepush(Q,(k+1,j)) \Commentadd to queue \EndIf\EndWhile\State𝐳=f(𝐁) \Comment f:T𝒵T maps matrix representation of tree to sequence of actions. \Statereturn 𝐳 \EndProcedure

3 Experimental Setup

3.1 Data

For English we use the Penn Treebank (Marcus et al., 1993, PTB) with splits and preprocessing from Dyer et al. (2016) which retains punctuation and replaces singleton words with Berkeley parser’s mapping rules, resulting in a vocabulary of 23,815 word types.1010 10 https://github.com/clab/rnng Notably this is much larger than the standard PTB LM setup from Mikolov et al. (2010) which uses 10K types.1111 11 Both versions of the PTB data can be obtained from http://demo.clab.cs.cmu.edu/cdyer/ptb-lm.tar.gz. Also different from the LM setup, we model each sentence separately instead of carrying information across sentence boundaries, as the RNNG is a generative model of sentences. Hence our perplexity numbers are not comparable to the PTB LM results (Melis et al., 2018; Merity et al., 2018; Yang et al., 2018).

Since the PTB is rather small, and since the URNNG does not require annotation, we also test our approach on a subset of the one billion word corpus (Chelba et al., 2013). We randomly sample 1M sentences for training and 2K sentences for validation/test, and limit the vocabulary to 30K word types. While still a subset of the full corpus (which has 30M sentences), this dataset is two orders of magnitude larger than PTB. Experiments on Chinese utilize version 5.1 of the Chinese Penn Treebank (CTB) (Xue et al., 2005), with the same splits as in Chen and Manning (2014). Singleton words are replaced with a single unk token, resulting in a vocabulary of 17,489 word types. {algorithm}[t] Calculating the tree entropy [qϕ(𝐳|𝐱)] {algorithmic}[1] \ProcedureEntropyβ \Commentβ from running Inside(s) \Fori:=1 to T \Commentinitialize entropy table \StateH[i,i]=0 \EndFor\Forl:=1 to T-1 \Commentspan length \Fori:=1 to T-l \Commentspan start \Statej=i+l \Commentspan end \Stateτ=u=ij-1β[i,u]β[u+1,j] \Foru:=i to j-1 \Statewu=(β[i,u]β[u+1,j])/τ \EndFor\StateH[i,j]=u=ij-1(H[i,u]+H[u+1,j] \State-logwu)wu \EndFor\EndFor\Statereturn H[1,T]\Commentreturn tree entropy [qϕ(𝐳|𝐱)] \EndProcedure

3.2 Training and Hyperparameters

The stack LSTM has two layers with input/hidden size equal to 650 and dropout of 0.5. The tree LSTM also has 650 units. The inference network uses a one-layer bidirectional LSTM with 256 hidden units, and the MLP (to produce span scores sij for ij) has a single hidden layer with a ReLU nonlinearity followed by layer normalization (Ba et al., 2016) and dropout of 0.5. We share word embeddings between the generative model and the inference network, and also tie weights between the input/output word embeddings (Press and Wolf, 2016).

Optimization of the model itself required standard techniques for avoiding posterior collapse in VAEs.1212 12 Posterior collapse in our context means that qϕ(𝐳|𝐱) always produced trivial (always left or right branching) trees. We warm-up the ELBO objective by linearly annealing (per batch) the weight on the conditional prior logpθ(𝐳|𝐱<𝐳) and the entropy [qϕ(𝐳|𝐱)] from 0 to 1 over the first two epochs (see equation (1) for definition of logpθ(𝐳|𝐱<𝐳)). This is analogous to KL-annealing in VAEs with continuous latent variables (Bowman et al., 2016; Sønderby et al., 2016). We train for 18 epochs (enough for convergence for all models) with a batch size of 16 and K=8 samples for the Monte Carlo gradient estimators. The generative model is optimized with SGD with learning rate equal to 1, except for the affine layer that produces a distribution over the actions, which has learning rate 0.1. Gradients of the generative model are clipped at 5. The inference network is optimized with Adam (Kingma and Ba, 2015) with learning rate 0.0001, β1=0.9,β2=0.999, and gradient clipping at 1. As Adam converges significantly faster than SGD (even with a much lower learning rate), we stop training the inference network after the first two epochs. Initial model parameters are sampled from 𝒰[-0.1,0.1]. The learning rate starts decaying by a factor of 2 each epoch after the first epoch at which validation performance does not improve, but this learning rate decay is not triggered for the first eight epochs to ensure adequate training. We use the same hyperparameters/training setup for both PTB and CTB. For experiments on (the subset of) the one billion word corpus, we use a smaller dropout rate of 0.1. The baseline RNNLM also uses the smaller dropout rate.

All models are trained with an end-of-sentence token, but for perplexity calculation these tokens are not counted to be comparable to prior work (Dyer et al., 2016; Kuncoro et al., 2017; Buys and Blunsom, 2018). To be more precise, the inference network does not make use of the end-of-sentence token to produce parse trees, but the generative model is trained to generate the end-of-sentence token after the final reduce operation.

3.3 Baselines

We compare the unsupervised RNNG (URNNG) against several baselines: (1) RNNLM, a standard RNN language model whose size is the same as URNNG’s stack LSTM; (2) Parsing Reading Predict Network (PRPN) (Shen et al., 2018), a neural language model that uses gated attention layers to embed soft tree-like structures into a neural network (and among the current state-of-the-art in grammar induction from words on the full corpus); (3) RNNG with trivial trees (left branching, right branching, random); (4) supervised RNNG trained on unlabeled, binarized gold trees.1313 13 We use right branching binarization—Matsuzaki et al. (2005) find that differences between various binarization schemes have marginal impact. Our supervised RNNG therefore differs the original RNNG, which trains on non-binarized trees and does not ignore constituent labels. Note that the supervised RNNG also trains a discriminative parser qϕ(𝐳|𝐱) (alongside the generative model pθ(𝐱,𝐳)) in order to sample parse forests for perplexity evaluation (i.e. importance sampling). This discriminative parser has the same architecture as URNNG’s inference network. For all models, we perform early stopping based on validation perplexity.

4 Results and Discussion

PTB CTB
Model PPL F1 PPL F1
RNNLM 93.2 201.3
PRPN (default) 126.2 32.9 290.9 32.9
PRPN (tuned) 96.7 41.2 216.0 36.1
Left Branching Trees 100.9 10.3 223.6 12.4
Right Branching Trees 93.3 34.8 203.5 20.6
Random Trees 113.2 17.0 209.1 17.4
URNNG 90.6 40.7 195.7 29.1
RNNG 88.7 68.1 193.1 52.3
RNNG URNNG 85.9 67.7 181.1 51.9
Oracle Binary Trees 82.5 88.6
Table 1: Language modeling perplexity (PPL) and grammar induction F1 scores on English (PTB) and Chinese (CTB) for the different models. We separate results for those that make do not make use of annotated data (top) versus those that do (mid). Note that our PTB setup from Dyer et al. (2016) differs considerably from the usual language modeling setup (Mikolov et al., 2010) since we model each sentence independently and use a much larger vocabulary (see creftype 3.1).

4.1 Language Modeling

Table 1 shows perplexity for the different models on PTB/CTB. As a language model URNNG outperforms an RNNLM and is competitive with the supervised RNNG.1414 14 For RNNG and URNNG we estimate the log marginal likelihood (and hence, perplexity) with K=1000 importance-weighted samples, logpθ(𝐱)log(1Kk=1Klogp(𝐱,𝐳(k))qϕ(𝐳(k)|𝐱)). During evaluation only, we also flatten qϕ(𝐳|𝐱) by dividing span scores sij by a temperature term 2.0 before feeding it to the CRF. The left branching baseline performs poorly, implying that the strong performance of URNNG/RNNG is not simply due to the additional depth afforded by the tree LSTM composition function (a left branching tree, which always performs reduce when possible, is the “deepest” model). The right branching baseline is essentially equivalent to an RNNLM and hence performs similarly. We found PRPN with default hyperparameters (which obtains a perplexity of 62.0 in the PTB setup from Mikolov et al. (2010)) to not perform well, but tuning hyperparameters improves performance.1515 15 Using the code from https://github.com/yikangshen/PRPN, we tuned model size, initialization, dropout, learning rate, and use of batch normalization. The supervised RNNG performs well as a language model, despite being trained on the joint (rather than marginal) likelihood objective.1616 16 RNNG is trained to maximize logpθ(𝐱,𝐳) while URNNG is trained to maximize (a lower bound on) the language modeling objective logpθ(𝐱). This indicates that explicit modeling of syntax helps generalization even with richly-parameterized neural models. Encouraged by these observations, we also experiment with a hybrid approach where we train a supervised RNNG first and continue fine-tuning the model (including the inference network) on the URNNG objective (RNNG URNNG in Table 1).1717 17 We fine-tune for 10 epochs and use a smaller learning rate of 0.1 for the generative model. This approach results in nontrivial perplexity improvements, and suggests that it is potentially possible to improve language models with supervision on parsed data.

Figure 2: Perplexity of the different models grouped by sentence length on PTB.

In Figure 2 we show perplexity by sentence length. We find that a standard language model (RNNLM) is better at modeling short sentences, but underperforms models that explicitly take into account structure (RNNG/URNNG) when the sentence length is greater than 10. Table 2 (top) compares our results against prior work on this version of the PTB, and Table 2 (bottom) shows the results on a 1M sentence subset of the one billion word corpus, which is two orders of magnitude larger than PTB. On this larger dataset URNNG still improves upon the RNNLM. We also trained an RNNG (and RNNG URNNG) on this dataset by parsing the training set with the self-attentive parser from Kitaev and Klein (2018).1818 18 To parse the training set we use the benepar_en2 model from https://github.com/nikitakit/self-attentive-parser, which obtains an F1 score of 95.17 on the PTB test set. These models improve upon the RNNLM but not the URNNG, potentially highlighting the limitations of using predicted trees for supervising RNNGs.

PTB PPL
KN 5-gram (Dyer et al., 2016) 169.3
RNNLM (Dyer et al., 2016) 113.4
Original RNNG (Dyer et al., 2016) 102.4
Stack-only RNNG (Kuncoro et al., 2017) 101.2
Gated-Attention RNNG (Kuncoro et al., 2017) 100.9
Generative Dep. Parser (Buys and Blunsom, 2015) 138.6
RNNLM (Buys and Blunsom, 2018) 100.7
Sup. Syntactic NLM (Buys and Blunsom, 2018) 107.6
Unsup. Syntactic NLM (Buys and Blunsom, 2018) 125.2
PRPN (Shen et al., 2018) 96.7
This work:
   RNNLM 93.2
   URNNG 90.6
   RNNG 88.7
   RNNG URNNG 85.9
1M Sentences PPL
PRPN (Shen et al., 2018) 77.7
RNNLM 77.4
URNNG 71.8
RNNG 72.9
RNNG URNNG 72.0
Table 2: (Top) Comparison of this work as a language model against prior works on sentence-level PTB with preprocessing from Dyer et al. (2016). Note that previous versions of RNNG differ from ours in terms of parameterization and model size. (Bottom) Results on a subset (1M sentences) of the one billion word corpus. PRPN is the model from Shen et al. (2018), whose hyperparameters were tuned by us. RNNG is trained on predicted parse trees from the self-attentive parser from Kitaev and Klein (2018).

4.2 Grammar Induction

Table 1 also shows the F1 scores for grammar induction. Note that we induce latent trees directly from words on the full dataset.1919 19 Past work on grammar induction usually train/evaluate on short sentences and also assume access to gold POS tags (Klein and Manning, 2002; Smith and Eisner, 2004; Bod, 2006). However more recent works train directly on words (Jin et al., 2018; Shen et al., 2018; Drozdov et al., 2019). For RNNG/URNNG we obtain the highest scoring tree from qϕ(𝐳|𝐱) through the Viterbi inside (i.e. CKY) algorithm.2020 20 Alternatively, we could estimate argmax𝐳pθ(𝐳|𝐱) by sampling parse trees from qϕ(𝐳|𝐱) and using pθ(𝐱,𝐳) to rerank the output, as in Dyer et al. (2016). We calculate unlabeled F1 using evalb, which ignores punctuation and discards trivial spans (width-one and sentence spans).2121 21 Available at https://nlp.cs.nyu.edu/evalb/. We evaluate with COLLINS.prm parameter file and LABELED option equal to 0. We observe that the setup for grammar induction varies widely across different papers: lexicalized vs. unlexicalized; use of punctuation vs. not; separation of train/test sets; counting sentence-level spans for evaluation vs. ignoring them; use of additional data; length cutoff for training/evaluation; corpus-level F1 vs. sentence-level F1; and, more. In our survey of twenty or so papers, almost no two papers were identical in their setup. Such variation makes it difficult to meaningfully compare models across papers. Hence, we report grammar induction results mainly for the models and baselines considered in the present work. Since we compare F1 against the original, non-binarized trees (per convention), F1 scores of models using oracle binarized trees constitute the upper bounds.

We confirm the replication study of Htut et al. (2018) and find that PRPN is a strong model for grammar induction. URNNG performs on par with PRPN on English but PRPN does better on Chinese; both outperform right branching baselines. Table 3 further analyzes the learned trees and shows the F1 score of URNNG trees against other trees (left), and the recall of URNNG/PRPN trees against ground truth constituents (right). We find that trees induced by URNNG and PRPN are quite different; URNNG is more sensitive to SBAR and VP, while PRPN is better at identifying NP. While left as future work, this naturally suggests a hybrid approach wherein the intersection of constituents from URNNG and PRPN is used to create a corpus of partially annotated trees, which can be used to guide another model, e.g. via posterior regularization (Ganchev et al., 2010) or semi-supervision (Hwa, 1999). Finally, Table 4 compares our results using the same evaluation setup as in Drozdov et al. (2019), which differs considerably from our setup.

Tree PTB CTB
Gold 40.7 29.1
Left 9.2 8.4
Right 68.3 51.2
Self 92.3 87.3
RNNG 55.4 47.1
PRPN 41.0 47.2
Label URNNG PRPN
SBAR 74.8% 28.9%
NP 39.5% 63.9%
VP 76.6% 27.3%
PP 55.8% 55.1%
ADJP 33.9% 42.5%
ADVP 50.4% 45.1%
Table 3: (Left) F1 scores of URNNG against other trees. “Self” refers to another URNNG trained with a different random seed. (Right) Recall of constituents by label for URNNG and PRPN. Recall for a particular label is the fraction of ground truth constituents of that label that were identified by the model (as in Htut et al. (2018)).
F1 +PP
PRPN-UP 39.8 45.4
PRPN-LM 42.8 42.4
ON-LSTM (Shen et al., 2019) 49.4 -
DIORA (Drozdov et al., 2019) 49.6 56.2
PRPN (tuned) 49.0 49.9
URNNG 52.4 52.4
Table 4: PTB F1 scores using the same evaluation setup as Drozdov et al. (2019), which evaluates against binarized trees, counts punctuation and trivial spans, and uses sentence-level F1. +PP indicates a post-processing heuristic which directly attaches trailing punctuation to the root. This does not change URNNG results since it learns to do so anyway. Results with are copied from Table 1 of Drozdov et al. (2019).

4.3 Distributional Metrics

Table 5 shows some standard metrics related to the learned generative model/inference network. The “reconstruction” perplexity based on 𝔼qϕ(𝐳|𝐱)[logpθ(𝐱|𝐳)] is much lower than regular perplexity, and further, the Kullback-Leibler divergence between the conditional prior and the variational posterior, given by

𝔼qϕ(𝐳|𝐱)[logqϕ(𝐳|𝐱)pθ(𝐳|𝐱<𝐳)],

is highly nonzero. (See equation (1) for definitions of logpθ(𝐱|𝐳) and logpθ(𝐳|𝐱<𝐳)). This indicates that the latent space is being used in a meaningful way and that there is no posterior collapse (Bowman et al., 2016). As expected, the entropy of the variational posterior is much lower than the entropy of the conditional prior, but there is still some uncertainty in the posterior.

PTB CTB
RNNG URNNG RNNG URNNG
PPL 88.7 90.6 193.1 195.7
Recon. PPL 74.6 73.4 183.4 151.9
KL 7.10 6.13 11.11 8.91
Prior Entropy 7.65 9.61 9.48 15.13
Post. Entropy 1.56 2.28 6.23 5.75
Unif. Entropy 26.07 26.07 30.17 30.17
Table 5: Metrics related to the generative model/inference network for RNNG/URNNG. For the supervised RNNG we take the “inference network” to be the discriminative parser trained alongside the generative model (see creftype 3.3). Recon. PPL is the reconstruction perplexity based on 𝔼qϕ(𝐳|𝐱)[logpθ(𝐱|𝐳)], and KL is the Kullback-Leibler divergence. Prior entropy is the entropy of the conditional prior pθ(𝐳|𝐱<𝐳), and uniform entropy is the entropy of the uniform distribution over all binary trees. The KL/entropy metrics are averaged across sentences.

4.4 Syntactic Evaluation

We perform a syntactic evaluation of the different models based on the setup from Marvin and Linzen (2018): the model is given two minimally different sentences, one grammatical and one ungrammatical, and must identify the grammatical sentence by assigning it higher probability.2222 22 We modify the publicly available dataset from https://github.com/BeckyMarvin/LM_syneval to only keep sentence pairs that did not have any unknown words with respect to our vocabulary, resulting in 80K sentence pairs for evaluation. Further, we train on a much smaller corpus, and hence our results are not directly comparable. Table 6 shows the accuracy results. Overall the supervised RNNG significantly outperforms the other models, indicating opportunities for further work in unsupervised modeling. While the URNNG does slightly outperform an RNNLM, the distribution of errors made from both models are similar, and thus it is not clear whether the outperformance is simply due to better perplexity or learning different structural biases.

4.5 Limitations

There are several limitations to our approach. For one, the URNNG takes considerably more time/memory to train than a standard language model due to the O(T3) dynamic program in the inference network, multiple samples to obtain low-variance gradient estimators, and dynamic computation graphs that make efficient batching nontrivial.2323 23 The main time bottleneck is the dynamic compution graph, since the dynamic programming algorithm can be batched (however the latter is a significant memory bottleneck). We manually batch the shift and reduce operation as much as possible, though recent work on auto-batching (Neubig et al., 2017) could potentially make this easier/faster. The model is sensitive to hyperparameters and required various optimization strategies (e.g. separate optimizers for the inference network and the generative model) to avoid posterior collapse. Finally, the URNNG also seemed to rely heavily on punctuation to identify constituents and we were unable to improve upon a right-branching baseline when training the URNNG on a version of PTB where punctuation is removed.2424 24 Many prior works that induce trees directly from words often employ additional heuristics based on punctuation (Seginer, 2007; Ponvert et al., 2011; Spitkovsky et al., 2013; Parikh et al., 2014), as punctuation (e.g. comma) is usually a reliable signal for start/end of constituent spans. The URNNG still has to learn to rely on punctuation, similar to recent works such as depth-bounded PCFGs (Jin et al., 2018) and DIORA (Drozdov et al., 2019). In contrast, PRPN (Shen et al., 2018) and Ordered Neurons (Shen et al., 2019) induce trees by directly training on corpus without punctuation. We also reiterate that punctuation is used during training but ignored during evaluation (except in Table 4).

RNNLM PRPN RNNG URNNG
PPL 93.2 96.7 88.7 90.6
Overall 62.5% 61.9% 69.3% 64.6%
Subj. 63.5% 63.7% 89.4% 67.2%
Obj. Rel. 62.6% 61.0% 67.6% 65.7%
Refl. 60.7% 68.8% 57.3% 60.5%
NPI 58.7% 39.5% 46.8% 55.0%
Table 6: Syntactic evaluation based on the setup from Marvin and Linzen (2018). Subj. is subject-verb agreement in sentential complement, across prepositional phrase/subjective relative clause, and VP coordination; Obj. Rel. refers to subject-verb agreement in/across an objective relative clause; Refl. refers to reflexive pronoun agreement with antecedent; NPI is negative polarity items.

5 Related Work

There has been much work on incorporating tree structures into deep models for syntax-aware language modeling, both for unconditional (Emami and Jelinek, 2005; Buys and Blunsom, 2015; Dyer et al., 2016) and conditional (Yin and Neubig, 2017; Alvarez-Melis and Jaakkola, 2017; Rabinovich et al., 2017; Aharoni and Goldberg, 2017; Eriguchi et al., 2017; Wang et al., 2018; Gu¯ et al., 2018) cases. These approaches generally rely on annotated parse trees during training and maximizes the joint likelihood of sentence-tree pairs. Prior work on combining language modeling and unsupervised tree learning typically embed soft, tree-like structures as hidden layers of a deep network (Cho et al., 2014; Chung et al., 2017; Shen et al., 2018, 2019). In contrast, Buys and Blunsom (2018) make Markov assumptions and perform exact marginalization over latent dependency trees. Our work is also related to the recent line of work on learning latent trees as part of a deep model through supervision on other tasks, typically via differentiable structured hidden layers (Kim et al., 2017; Bradbury and Socher, 2017; Liu and Lapata, 2018; Tran and Bisk, 2018; Peng et al., 2018; Niculae et al., 2018; Liu et al., 2018), policy gradient-based approaches (Yogatama et al., 2017; Williams et al., 2018; Havrylov et al., 2019), or differentiable relaxations (Choi et al., 2018; Maillard and Clark, 2018).

The variational approximation uses amortized inference (Kingma and Welling, 2014; Mnih and Gregor, 2014; Rezende et al., 2014), in which an inference network is used to obtain the variational posterior for each observed 𝐱. Since our inference network is structured (i.e., a CRF), it is also related to CRF autoencoders (Ammar et al., 2014) and structured VAEs (Johnson et al., 2016; Krishnan et al., 2017), which have been used previously for unsupervised (Cai et al., 2017; Drozdov et al., 2019; Li et al., 2019) and semi-supervised (Yin et al., 2018; Corro and Titov, 2019) parsing.

6 Conclusion

It is an open question as to whether explicit modeling of syntax significantly helps neural models. Strubell et al. (2018) find that supervising intermediate attention layers with syntactic heads improves semantic role labeling, while Shi et al. (2018) observe that for text classification, syntactic trees only have marginal impact. Our work suggests that at least for language modeling, incorporating syntax either via explicit supervision or as latent variables does provide useful inductive biases and improves performance.

Finally, in modeling child language acquisition, the complex interaction of the parser and the grammatical knowledge being acquired is the object of much investigation (Trueswell and Gleitman, 2007); our work shows that apparently grammatical constraints can emerge from the interaction of a constrained parser and a more general grammar learner, which is an intriguing but underexplored hypothesis for explaining human linguistic biases.

Acknowledgments

We thank the members of the DeepMind language team for helpful feedback. YK is supported by a Google Fellowship. AR is supported by NSF Career 1845664.

References

  • Aharoni and Goldberg (2017) Roee Aharoni and Yoav Goldberg. 2017. Towards String-to-Tree Neural Machine Translation. In Proceedings of ACL.
  • Alvarez-Melis and Jaakkola (2017) David Alvarez-Melis and Tommi S. Jaakkola. 2017. Tree-structured Decoding with Doubly-Recurrent Neural Networks. In Proceedings of ICLR.
  • Ammar et al. (2014) Waleed Ammar, Chris Dyer, and Noah A. Smith. 2014. Conditional Random Field Autoencoders for Unsupervised Structured Prediction. In Proceedings of NIPS.
  • Ba et al. (2016) Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. 2016. Layer Normalization. In Proceedings of NIPS.
  • Baker (1979) James K. Baker. 1979. Trainable Grammars for Speech Recognition. In Proceedings of the Spring Conference of the Acoustical Society of America.
  • Bod (2006) Rens Bod. 2006. An All-Subtrees Approach to Unsupervised Parsing. In Proceedings of ACL.
  • Bowman et al. (2016) Samuel R. Bowman, Luke Vilnis, Oriol Vinyal, Andrew M. Dai, Rafal Jozefowicz, and Samy Bengio. 2016. Generating Sentences from a Continuous Space. In Proceedings of CoNLL.
  • Bradbury and Socher (2017) James Bradbury and Richard Socher. 2017. Towards Neural Machine Translation with Latent Tree Attention. In Proceedings of the 2nd Workshop on Structured Prediction for Natural Language Processing.
  • Buys and Blunsom (2015) Jan Buys and Phil Blunsom. 2015. Generative Incremental Dependency Parsing with Neural Networks. In Proceedings of ACL.
  • Buys and Blunsom (2018) Jan Buys and Phil Blunsom. 2018. Neural Syntactic Generative Models with Exact Marginalization. In Proceedings of NAACL.
  • Cai et al. (2017) Jiong Cai, Yong Jiang, and Kewei Tu. 2017. CRF Autoencoder for Unsupervised Dependency Parsing. In Proceedings of EMNLP.
  • Chelba et al. (2013) Ciprian Chelba, Tomas Mikolov, Mike Schuster, Qi Ge, Thorsten Brants, Phillipp Koehn, and Tony Robinson. 2013. One Billion Word Benchmark for Measuring Progress in Statistical Language Modeling. arXiv:1312.3005.
  • Chen and Manning (2014) Danqi Chen and Christopher D. Manning. 2014. A Fast and Accurate Dependency Parser using Neural Networks. In Proceedings of EMNLP.
  • Cho et al. (2014) Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. 2014. On the Properties of Neural Machine Translation: Encoder-Decoder Approaches. In Proceedings of Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation.
  • Choi et al. (2018) Jihun Choi, Kang Min Yoo, and Sang goo Lee. 2018. Learning to Compose Task-Specific Tree Structures. In Proceedings of AAAI.
  • Chung et al. (2017) Junyoung Chung, Sungjin Ahn, and Yoshua Bengio. 2017. Hierarchical Multiscale Recurrent Neural Networks. In Proceedings of ICLR.
  • Corro and Titov (2019) Caio Corro and Ivan Titov. 2019. Differentiable Perturb-and-Parse: Semi-Supervised Parsing with a Structured Variational Autoencoder. In Proceedings of ICLR.
  • Cremer et al. (2018) Chris Cremer, Xuechen Li, and David Duvenaud. 2018. Inference Suboptimality in Variational Autoencoders. In Proceedings of ICML.
  • Deng et al. (2018) Yuntian Deng, Yoon Kim, Justin Chiu, Demi Guo, and Alexander M. Rush. 2018. Latent Alignment and Variational Attention. In Proceedings of NIPS.
  • Drozdov et al. (2019) Andrew Drozdov, Patrick Verga, Mohit Yadev, Mohit Iyyer, and Andrew McCallum. 2019. Unsupervised Latent Tree Induction with Deep Inside-Outside Recursive Auto-Encoders. In Proceedings of NAACL.
  • Durrett and Klein (2015) Greg Durrett and Dan Klein. 2015. Neural CRF Parsing. In Proceedings of ACL.
  • Dyer et al. (2015) Chris Dyer, Miguel Ballesteros, Wang Ling, Austin Matthews, and Noah A. Smith. 2015. Transition-Based Dependency Parsing with Stack Long Short-Term Memory. In Proceedings of ACL.
  • Dyer et al. (2016) Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith. 2016. Recurrent Neural Network Grammars. In Proceedings of NAACL.
  • Emami and Jelinek (2005) Ahmad Emami and Frederick Jelinek. 2005. A Neural Syntactic Language Model. Machine Learning, 60:195–227.
  • Eriguchi et al. (2017) Akiko Eriguchi, Yoshimasa Tsuruoka, and Kyunghyun Cho. 2017. Learning to Parse and Translate Improves Neural Machine Translation. In Proceedings of ACL.
  • Finkel et al. (2008) Jenny Rose Finkel, Alex Kleeman, and Christopher D. Manning. 2008. Efficient, Feature-based, Conditional Random Field Parsing. In Proceedings of ACL.
  • Finkel et al. (2006) Jenny Rose Finkel, Christopher D. Manning, and Andrew Y. Ng. 2006. Solving the Problem of Cascading Errors: Approximate Bayesian Inference for Linguistic Annotation Pipelines. In Proceedings of EMNLP.
  • Ganchev et al. (2010) Kuzman Ganchev, João Graça, Jennifer Gillenwater, and Ben Taskar. 2010. Posterior Regularization for Structured Latent Variable Models. Journal of Machine Learning Research, 11:2001–2049.
  • Glynn (1987) Peter Glynn. 1987. Likelihood Ratio Gradient Estimation: An Overview. In Proceedings of Winter Simulation Conference.
  • Goodman (1998) Joshua Goodman. 1998. Parsing Inside-Out. PhD thesis, Harvard University.
  • Gu¯ et al. (2018) Jetic Gu¯, Hassan S. Shavarani, and Anoop Sarkar. 2018. Top-down Tree Structured Decoding with Syntactic Connections for Neural Machine Translation and Parsing. In Proceedings of EMNLP.
  • Hale et al. (2018) John Hale, Chris Dyer, Adhiguna Kuncoro, and Jonathan R. Brennan. 2018. Finding Syntax in Human Encephalography with Beam Search. In Proceedings of ACL.
  • Havrylov et al. (2019) Serhii Havrylov, Germán Kruszewski, and Armand Joulin. 2019. Cooperative Learning of Disjoint Syntax and Semantics. In Proceedings of NAACL.
  • Htut et al. (2018) Phu Mon Htut, Kyunghyun Cho, and Samuel R. Bowman. 2018. Grammar Induction with Neural Language Models: An Unusual Replication. In Proceedings of EMNLP.
  • Hwa (1999) Rebecca Hwa. 1999. Supervised Grammar Induction Using Training Data with Limited Constituent Information. In Proceedings of ACL.
  • Hwa (2000) Rebecca Hwa. 2000. Sample Selection for Statistical Grammar Induction. In Proceedings of EMNLP.
  • Jin et al. (2018) Lifeng Jin, Finale Doshi-Velez, Timothy Miller, William Schuler, and Lane Schwartz. 2018. Unsupervised Grammar Induction with Depth-bounded PCFG. In Proceedings of TACL.
  • Johnson et al. (2007) Mark Johnson, Thomas L. Griffiths, and Sharon Goldwater. 2007. Bayesian Inference for PCFGs via Markov chain Monte Carlo. In Proceedings of NAACL.
  • Johnson et al. (2016) Matthew Johnson, David K. Duvenaud, Alex Wiltschko, Ryan P. Adams, and Sandeep R. Datta. 2016. Composing Graphical Models with Neural Networks for Structured Representations and Fast Inference. In Proceedings of NIPS.
  • Kawakami et al. (2018) Kazuya Kawakami, Chris Dyer, and Phil Blunsom. 2018. Unsupervised Word Discovery with Segmental Neural Language Models. arXiv:1811.09353.
  • Kim et al. (2017) Yoon Kim, Carl Denton, Luong Hoang, and Alexander M. Rush. 2017. Structured Attention Networks. In Proceedings of ICLR.
  • Kingma and Ba (2015) Diederik P. Kingma and Jimmy Ba. 2015. Adam: A Method for Stochastic Optimization. In Proceedings of ICLR.
  • Kingma and Welling (2014) Diederik P. Kingma and Max Welling. 2014. Auto-Encoding Variational Bayes. In Proceedings of ICLR.
  • Kitaev and Klein (2018) Nikita Kitaev and Dan Klein. 2018. Constituency Parsing with a Self-Attentive Encoder. In Proceedings of ACL.
  • Klein and Manning (2002) Dan Klein and Christopher Manning. 2002. A Generative Constituent-Context Model for Improved Grammar Induction. In Proceedings of ACL.
  • Krishnan et al. (2017) Rahul G. Krishnan, Uri Shalit, and David Sontag. 2017. Structured Inference Networks for Nonlinear State Space Models. In Proceedings of AAAI.
  • Kuncoro et al. (2017) Adhiguna Kuncoro, Miguel Ballesteros, Lingpeng Kong, Chris Dyer, Graham Neubig, and Noah A. Smith. 2017. What Do Recurrent Neural Network Grammars Learn About Syntax? In Proceedings of EACL.
  • Kuncoro et al. (2018) Adhiguna Kuncoro, Chris Dyer, John Hale, Dani Yogatama, Stephen Clark, and Phil Blunsom. 2018. LSTMs Can Learn Syntax-Sensitive Dependencies Well, But Modeling Structure Makes Them Better. In Proceedings of ACL.
  • Li et al. (2019) Bowen Li, Jianpeng Cheng, Yang Liu, and Frank Keller. 2019. Dependency Grammar Induction with a Neural Variational Transition-based Parser. In Proceedings of AAAI.
  • Li and Eisner (2009) Zhifei Li and Jason Eisner. 2009. First- and Second-Order Expectation Semirings with Applications to Minimum-Risk Training on Translation Forests. In Proceedings of EMNLP.
  • Liu et al. (2018) Yang Liu, Matt Gardner, and Mirella Lapata. 2018. Structured Alignment Networks for Matching Sentences. In Proceedings of EMNLP.
  • Liu and Lapata (2018) Yang Liu and Mirella Lapata. 2018. Learning Structured Text Representations. In Proceedings of TACL.
  • Maillard and Clark (2018) Jean Maillard and Stephen Clark. 2018. Latent Tree Learning with Differentiable Parsers: Shift-Reduce Parsing and Chart Parsing. In Proceedings of the Workshop on the Relevance of Linguistic Structure in Neural Architectures for NLP.
  • Marcus et al. (1993) Mitchell P. Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. 1993. Building a Large Annotated Corpus of English: The Penn Treebank. Computational Linguistics, 19:313–330.
  • Marvin and Linzen (2018) Rebecca Marvin and Tal Linzen. 2018. Targeted Syntactic Evaluation of Language Models. In Proceedings of EMNLP.
  • Matsuzaki et al. (2005) Takuya Matsuzaki, Yusuke Miyao, and Junichi Tsujii. 2005. Probabilistic CFG with Latent Annotations. In Proceedings of ACL.
  • Melis et al. (2018) Gábor Melis, Chris Dyer, and Phil Blunsom. 2018. On the State of the Art of Evaluation in Neural Language Models. In Proceedings of ICLR.
  • Merity et al. (2018) Stephen Merity, Nitish Shirish Keskar, and Richard Socher. 2018. Regularizing and Optimizing LSTM Language Models. In Proceedings of ICLR.
  • Mikolov et al. (2010) Tomas Mikolov, Martin Karafiat, Lukas Burget, Jan Cernocky, and Sanjeev Khudanpur. 2010. Recurrent Neural Network Based Language Model. In Proceedings of INTERSPEECH.
  • Mnih and Gregor (2014) Andriy Mnih and Karol Gregor. 2014. Neural variational inference and learning in belief networks. In Proceedings of ICML.
  • Mnih and Rezende (2016) Andriy Mnih and Danilo J. Rezende. 2016. Variational Inference for Monte Carlo Objectives. In Proceedings of ICML.
  • Naseem et al. (2010) Tahira Naseem, Harr Chen, Regina Barzilay, and Mark Johnson. 2010. Using Universal Linguistic Knowledge to Guide Grammar Induction. In Proceedings of EMNLP.
  • Neubig et al. (2017) Graham Neubig, Yoav Goldberg, and Chris Dyer. 2017. On-the-fly Operation Batching in Dynamic Computation Graphs. arXiv:1705.07860.
  • Niculae et al. (2018) Vlad Niculae, André F. T. Martins, and Claire Cardie. 2018. Towards Dynamic Computation Graphs via Sparse Latent Structure. In Proceedings of EMNLP.
  • Parikh et al. (2014) Ankur P. Parikh, Shay B. Cohen, and Eric P. Xing. 2014. Spectral Unsupervised Parsing with Additive Tree Metrics. In Proceedings of ACL.
  • Peng et al. (2018) Hao Peng, Sam Thomson, and Noah A. Smith. 2018. Backpropagating through Structured Argmax using a SPIGOT. In Proceedings of ACL.
  • Ponvert et al. (2011) Elis Ponvert, Jason Baldridge, and Katrin Erk. 2011. Simpled Unsupervised Grammar Induction from Raw Text with Cascaded Finite State Methods. In Proceedings of ACL.
  • Press and Wolf (2016) Ofir Press and Lior Wolf. 2016. Using the Output Embedding to Improve Language Models. In Proceedings of EACL.
  • Rabinovich et al. (2017) Maxim Rabinovich, Mitchell Stern, and Dan Klein. 2017. Abstract Syntax Networks for Code Generation and Semantic Parsing. In Proceedings of ACL.
  • Rezende et al. (2014) Danilo J. Rezende, Shakir Mohamed, and Daan Wierstra. 2014. Stochastic Backpropagation and Approximate Inference in Deep Generative Models. In Proceedings of ICML.
  • Seginer (2007) Yoav Seginer. 2007. Fast Unsupervised Incremental Parsing. In Proceedings of ACL.
  • Shen et al. (2018) Yikang Shen, Zhouhan Lin, Chin-Wei Huang, and Aaron Courville. 2018. Neural Language Modeling by Jointly Learning Syntax and Lexicon. In Proceedings of ICLR.
  • Shen et al. (2019) Yikang Shen, Shawn Tan, Alessandro Sordoni, and Aaron Courville. 2019. Ordered Neurons: Integrating Tree Structures into Recurrent Neural Networks. In Proceedings of ICLR.
  • Shi et al. (2018) Haoyue Shi, Hao Zhou, Jiaze Chen, and Lei Li. 2018. On Tree-based Neural Sentence Modeling. In Proceedings of EMNLP.
  • Smith and Eisner (2004) Noah A. Smith and Jason Eisner. 2004. Annealing Techniques for Unsupervised Statistical Language Learning. In Proceedings of ACL.
  • Sønderby et al. (2016) Casper Kaae Sønderby, Tapani Raiko, Lars Maaløe, Søren Kaae Sønderby, and Ole Winther. 2016. Ladder Variational Autoencoders. In Proceedings of NIPS.
  • Spitkovsky et al. (2013) Valentin I. Spitkovsky, Hiyan Alshawi, and Daniel Jurafsky. 2013. Breaking Out of Local Optima with Count Transforms and Model Recombination: A Study in Grammar Induction. In Proceedings of EMNLP.
  • Stern et al. (2017) Mitchell Stern, Jacob Andreas, and Dan Klein. 2017. A Minimal Span-Based Neural Constituency Parser. In Proceedings of ACL.
  • Strubell et al. (2018) Emma Strubell, Patrick Verga, Daniel Andor, David Weiss, and Andrew McCallum. 2018. Linguistically-Informed Self-Attention for Semantic Role Labeling. In Proceedings of EMNLP.
  • Tai et al. (2015) Kai Sheng Tai, Richard Socher, and Christopher D. Manning. 2015. Improved Semantic Representations From Tree-Structured Long Short-Term Memory Networks. In Proceedings of ACL.
  • Tran and Bisk (2018) Ke Tran and Yonatan Bisk. 2018. Inducing Grammars with and for Neural Machine Translation. In Proceedings of the 2nd Workshop on Neural Machine Translation and Generation.
  • Trueswell and Gleitman (2007) John C. Trueswell and Lila R. Gleitman. 2007. Learning to Parse and its Implications for Language Acquisition. The Oxford Handbook of Psycholinguistics.
  • Wang and Chang (2016) Wenhui Wang and Baobao Chang. 2016. Graph-based Dependency Parsing with Bidirectional LSTM. In Proceedings of ACL.
  • Wang et al. (2018) Xinyi Wang, Hieu Pham, Pengcheng Yin, and Graham Neubig. 2018. A Tree-based Decoder for Neural Machine Translation. In Proceedings of EMNLP.
  • Williams et al. (2018) Adina Williams, Andrew Drozdov, and Samuel R. Bowman. 2018. Do Latent Tree Learning Models Identify Meaningful Structure in Sentences? In Proceedings of TACL.
  • Williams (1992) Ronald J. Williams. 1992. Simple Statistical Gradient-following Algorithms for Connectionist Reinforcement Learning. Machine Learning, 8.
  • Wiseman et al. (2018) Sam Wiseman, Stuart M. Shieber, and Alexander M. Rush. 2018. Learning Neural Templates for Text Generation. In Proceedings of EMNLP.
  • Xue et al. (2005) Naiwen Xue, Fei Xia, Fu dong Chiou, and Marta Palmer. 2005. The Penn Chinese Treebank: Phrase Structure Annotation of a Large Corpus. Natural Language Engineering, 11:207–238.
  • Yang et al. (2018) Zhilin Yang, Zihang Dai, Ruslan Salakhutdinov, and William W. Cohen. 2018. Breaking the Softmax Bottleneck: A High-Rank RNN Language Model. In Proceedings of ICLR.
  • Yin and Neubig (2017) Pengcheng Yin and Graham Neubig. 2017. A Syntactic Neural Model for General-Purpose Code Generation. In Proceedings of ACL.
  • Yin et al. (2018) Pengcheng Yin, Chunting Zhou, Junxian He, and Graham Neubig. 2018. StructVAE: Tree-structured Latent Variable Models for Semi-supervised Semantic Parsing. In Proceedings of ACL.
  • Yogatama et al. (2017) Dani Yogatama, Phil Blunsom, Chris Dyer, Edward Grefenstette, and Wang Ling. 2017. Learning to Compose Words into Sentences with Reinforcement Learning. In Proceedings of ICLR.
  • Zhu et al. (2015) Xiaodan Zhu, Parinaz Sobhani, and Hongyu Guo. 2015. Long Short-Term Memory Over Tree Structures. In Proceedings of ICML.