A Generative Model for Punctuation in Dependency Trees

Punctuation enriches the expressiveness of written language. When converting from spoken to written language, punctuation indicates pauses or pitches; expresses propositional attitude; and is conventionally associated with certain syntactic constructions such as apposition, parenthesis, quotation, and conjunction.

In this paper, we present a latent-variable model of punctuation usage, inspired by the rule-based approach to English punctuation of Nunberg (1990). Training our model on English data learns rules that are consistent with Nunberg’s hand-crafted rules. Our system is automatic, so we use it to obtain rules for Arabic, Chinese, Spanish, and Hindi as well.

Moreover, our rules are stochastic, which allows us to reason probabilistically about ambiguous or missing punctuation. Across the 5 languages, our model predicts surface punctuation better than baselines, as measured both by perplexity (§4) and by accuracy on a punctuation restoration task (§6.1). We also use our model to correct the punctuation of non-native writers of English (§6.2), and to maintain natural punctuation style when syntactically transforming English sentences (§6.3). In principle, our model could also be used within a generative parser, allowing the parser to evaluate whether a candidate tree truly explains the punctuation observed in the input sentence (§8).

In The Linguistics of Punctuation, Nunberg (1990) argues that punctuation (in English) is more than a visual counterpart of spoken-language prosody, but forms a linguistic system that involves “interactions of point indicators (i.e. commas, semicolons, colons, periods and dashes).” He proposes that much as in phonology (Chomsky and Halle, 1968), a grammar generates underlying punctuation which then transforms into the observed surface punctuation.

Consider generating a sentence from a syntactic grammar as follows:

Although the full tree is not depicted here, some of the constituents are indicated with brackets. In this underlying generated tree, each appositive NP is surrounded by commas. On the surface, however, the two adjacent commas after Pendragon will now be collapsed into one, and the final comma will be absorbed into the adjacent period. Furthermore, in American English, the typographic convention is to move the final punctuation inside the quotation marks. Thus a reader sees only this modified surface form of the sentence:

Note that these modifications are string transformations that do not see or change the tree. The resulting surface punctuation marks may be clues to the parse tree, but (contrary to NLP convention) they should not be included as nodes in the parse tree. Only the underlying marks play that role.

Pang et al. (2002) use question and exclamation marks as clues to sentiment. Similarly, quotation marks may be used to mark titles, quotations, reported speech, or dubious terminology (University of Chicago, 2010). Because of examples like this, methods for determining the similarity or meaning of syntax trees, such as a tree kernel (Agarwal et al., 2011) or a recursive neural network (Tai et al., 2015), should ideally be able to consider where the underlying punctuation marks attach.

Surface punctuation remains correlated with syntactic phrase structure. NLP systems for generating or editing text must be able to deploy surface punctuation as human writers do. Parsers and grammar induction systems benefit from the presence of surface punctuation marks (Jones, 1994; Spitkovsky et al., 2011). It is plausible that they could do better with a linguistically informed model that explains exactly why the surface punctuation appears where it does. Patterns of punctuation usage can also help identify the writer’s native language (Markov et al., 2018).

Work on syntax and parsing tends to treat punctuation as an afterthought rather than a phenomenon governed by its own linguistic principles. Treebank annotation guidelines for punctuation tend to adopt simple heuristics like “attach to the highest possible node that preserves projectivity” (Bies et al., 1995; Nivre et al., 2018).¹ Many dependency parsing works exclude punctuation from evaluation (Nivre et al., 2007b; Koo and Collins, 2010; Chen and Manning, 2014; Lei et al., 2014; Kiperwasser and Goldberg, 2016), although some others retain punctuation (Nivre et al., 2007a; Goldberg and Elhadad, 2010; Dozat and Manning, 2017).

In tasks such as word embedding induction (Mikolov et al., 2013; Pennington et al., 2014) and machine translation (Zens et al., 2002), punctuation marks are usually either removed or treated as ordinary words (Řehůřek and Sojka, 2010).

Yet to us, building a parse tree on a surface sentence seems as inappropriate as morphologically segmenting a surface word. In both cases, one should instead analyze the latent underlying form, jointly with recovering that form. For example, the proper segmentation of English hoping is not hop-ing but hope-ing (with underlying e), and the proper segmentation of stopping is neither stopp-ing nor stop-ping but stop-ing (with only one underlying p). Cotterell et al. (2015); Cotterell et al. (2016) get this right for morphology. We attempt to do the same for punctuation.

This third step may alter the sequence of punctuation tokens at each slot between words—for example, in §1, collapsing the double comma , , between Pendragon and who. u and x denote just the punctuation at the slots of ū and $x¯$ respectively, with u_i and x_i denoting the punctuation token sequences at the i^th slot. Thus, the transformation at the i^th slot is u_i↦x_i.

Since this model is generative, we could train it without any supervision to explain the observed surface string $x¯$⁠: maximize the likelihood $p(x¯)$ in (1), marginalizing out the possible T, T′ values.

From the tree T′, we can read off the sequence of underlying punctuation tokens u_i at each slot i between words. Namely, u_i concatenates the right punctemes of all constituents ending at i with the left punctemes of all constituents starting at i (as illustrated by the examples in §1 and Figure 1). The NoisyChannel model then transduces u_i to a surface token sequence x_i, for each i = 0, …, n independently (where n is the sentence length).

Much like Halle’s (1968) phonological grammar of English, Nunberg’s 1990 descriptive English punctuation grammar (Table 1) can be viewed computationally as a priority string rewriting system, or Markov algorithm (Markov, 1960; Caracciolo di Forino, 1968). The system begins with a token string u. At each step it selects the highest-priority local rewrite rule that can apply, and applies it as far left as possible. When no more rules can apply, the final state of the string is returned as x.

Markov algorithms are Turing complete. Fortunately, Johnson (1972) noted that in practice, phonological u ↦ x maps described in this formalism can usually be implemented with finite-state transducers (FSTs).

For computational simplicity, we will formulate our punctuation model as a probabilistic FST (PFST)—a locally normalized left-to-right rewrite model (Cotterell et al., 2014). The probabilities for each language must be learned, using gradient descent. Normally we expect most probabilities to be near 0 or 1, making the PFST nearly deterministic (i.e., close to a subsequential FST). However, permitting low-probability choices remains useful to account for typographical errors, dialectal differences, and free variation in the training corpus.

Our PFST generates a surface string, but the invertibility of FSTs will allow us to work backwards when analyzing a surface string (§3).

Instead of having rule priorities, we apply Nunberg-style rules within a 2-token window that slides over u in a single left-to-right pass (Figure 2). Conditioned on the current window contents ab, a single edit is selected stochastically: either ab ↦ ab (no change), ab ↦ b (left absorption), ab ↦ a (right absorption), or ab ↦ ba (transposition). Then the window slides rightward to cover the next input token, together with the token that is (now) to its left. a and b are always real tokens, never boundary symbols. ϕ specifies the conditional edit probabilities.⁵

These specific edit rules (like Nunberg’s) cannot insert new symbols, nor can they delete all of the underlying symbols. Thus, surface x_i is a good clue to u_i: all of its tokens must appear underlyingly, and if x_i = ϵ (the empty string) then u_i = ϵ.

The model can be directly implemented as a PFST (Appendix D4) using Cotterell et al.’s (2014) more general PFST construction.

Our single-pass formalism is less expressive than Nunberg’s. It greedily makes decisions based on at most one token of right context (“label bias”). It cannot rewrite ’”. ↦ .’” or “,. ↦ .’” because the . is encountered too late to percolate leftward; luckily, though, we can handle such English examples by sliding the window right-to-left instead of left-to-right. We treat the sliding direction as a language-specific parameter.⁶

The expectation $E[⋯]$ is over $T′∼p(⋅∣T,x)$⁠. This generalized expectation term provides posterior regularization (Mann and McCallum, 2010; Ganchev et al., 2010), by encouraging parameters that reconstruct trees T′ that use symmetric punctuation marks in a “typical” way. The function c(T′) counts the nodes in T′ whose punctemes contain “unmatched” symmetric punctuation tokens: for example, ) is “matched” only when it appears in a right puncteme with ( at the comparable position in the same constituent’s left puncteme. The precise definition is given in Appendix B.4

In our development experiments on English, the posterior regularization term was necessary to discover an aesthetically appealing theory of underlying punctuation. When we dropped this term (ξ = 0) and simply maximized the ordinary regularized likelihood, we found that the optimization problem was underconstrained: different training runs would arrive at different, rather arbitrary underlying punctemes. For example, one training run learned an Attach model that used underlying " to terminate sentences, along with a NoisyChannel model that absorbed the left quotation mark into the period. By encouraging the underlying punctuation to be symmetric, we broke the ties. We also tried making this a hard constraint (ξ = ∞), but then the model was unable to explain some of the training sentences at all, giving them probability of 0. For example, I went to the“ special place ” cannot be explained, because special place is not a constituent.⁸

In principle, working with the model (1) is straightforward, thanks to the closure properties of formal languages. Provided that p_syn can be encoded as a weighted CFG, it can be composed with the weighted tree transducer p_θ and the weighted FST p_ϕ to yield a new weighted CFG (similarly to Bar-Hillel et al., 1961; Nederhof and Satta, 2003). Under this new grammar, one can recover the optimal T,T′ for $x¯$ by dynamic programming, or sum over T,T′ by the inside algorithm to get the likelihood $p(x¯)$⁠. A similar approach was used by Levy (2008) with a different FST noisy channel.

In this paper we assume that T is observed, allowing us to work with (2). This cuts the computation time from O(n³) to O(n).⁹ Whereas the inside algorithm for (1) must consider O(n²) possible constituents of $x¯$ and O(n) ways of building each, our algorithm for (2) only needs to iterate over the O(n) true constituents of T and the 1 true way of building each. However, it must still consider the $|Wd|$ puncteme pairs for each constituent.

Given an input sentence $x¯$ of length n, our job is to sum over possible trees T′ that are consistent with T and $x¯$⁠, or to find the best such T′. This is roughly a lattice parsing problem—made easier by knowing T. However, the possible ū values are characterized not by a lattice but by a cyclic WFSA (as |u_i| is unbounded whenever |x_i| > 0).

For each slot 0 ≤ i ≤ n, transduce the surface punctuation string x_i by the inverted PFST for p_ϕ to obtain a weighted finite-state automaton (WFSA) that describes all possible underlying strings u_i.¹⁰ This WFSA accepts each possible u_i with weight p_ϕ(x_i∣u_i). If it has N_i states, we can represent it (Berstel and Reutenauer, 1988) with a family of sparse weight matrices $Mi(υ)∈RNi×Ni$⁠, whose element at row s and column t is the weight of the s → t arc labeled with υ, or 0 if there is no such arc. Additional vectors $λi,ρi∈RNi$ specify the initial and final weights. (λ_i is one-hot if the PFST has a single initial state, of weight 1.)

For any puncteme l (or r) in $V$⁠, we define M_i(l) = M_i(l₁)M_i(l₂)⋯M_i(l_|l|), a product over the 0 or more tokens in l. This gives the total weight of all s →^*t WFSA paths labeled with l.

The subprocedure in Algorithm 1 essentially extends this to obtain a new matrix $In(w)∈RNi×Nk$⁠, where the subtree rooted at w stretches from slot i to slot k. Its element In(w)_st gives the total weight of all extended paths in the ū WFSA from state s at slot i to state t at slot k. An extended path is defined by a choice of underlying punctemes at w and all its descendants. These punctemes determine an s-to-final path at i, then initial-to-final paths at i + 1 through k − 1, then an initial-to-t path at k. The weight of the extended path is the product of all the WFSA weights on these paths (which correspond to transition probabilities in p_ϕ PFST) times the probability of the choice of punctemes (from p_θ).

This inside algorithm computes quantities needed for training (§2.3). Useful variants arise via well-known methods for weighted derivation forests (Berstel and Reutenauer, 1988; Goodman, 1999; Li and Eisner, 2009; Eisner, 2016).

Specifically, to modify Algorithm 1 to maximize over T′ values (§§6.2–6.3) instead of summing over them, we switch to the derivation semiring (Goodman, 1999), as follows. Whereas In(w)_st used to store the total weight of all extended paths from state s at slot i to state t at slot j, now it will store the weight of the best such extended path. It will also store that extended path’s choice of underlying punctemes, in the form of a puncteme-annotated version of the subtree of T that is rooted at w. This is a potential subtree of T′.

What is DD′? For presentational purposes, it is convenient to represent a punctuated dependency tree as a bracketed string. For example, the underlying tree T′ in Figure 1 would be [ [“ Dale ”] means [“ [ river ] valley ”] ] where the words correspond to nodes of T. In this case, we can represent every D as a partial bracketed string and define DD′ by string concatenation. This presentation ensures that multiplication (7) is a complete and associative (though not commutative) operation, as in any semiring. As base cases, each real-valued element of M_i(l) or M_k(r) is now paired with the string [l or r] respectively, ¹¹and the real number 1 at line 10 is paired with the string w. The real-valued elements of the λ_i and ρ_i vectors and the 0 matrix at line 11 are paired with the empty string ϵ, as is the real number p at line 13.

In practice, the D strings that appear within the matrix M of Algorithm 1 will always represent complete punctuated trees. Thus, they can actually be represented in memory as such, and different trees may share subtrees for efficiency (using pointers). The product in line 10 constructs a matrix of trees with root w and differing sequences of left/right children, while the product in line 14 annotates those trees with punctemes l,r.

Having computed the objective (5), we find the gradient via automatic differentiation, and optimize $θ,ϕ$ via Adam (Kingma and Ba, 2014)—a variant of stochastic gradient decent—with learning rate 0.07, batchsize 5, sentence per epoch 400, and L2 regularization. (These hyperparameters, along with the regularization coefficients ς and ξ from (5), were tuned on dev data (§4) for each language respectively.) We train the punctuation model for 30 epochs. The initial NoisyChannel parameters (ϕ) are drawn from $N(0,1)$⁠, and the initial Attach parameters (θ) are drawn from $N(0,1)$ (with one minor exception described in Appendix A).

Throughout §§4–6, we will examine the punctuation model on a subset of the Universal Dependencies (UD) version 1.4 (Nivre et al., 2016)—a collection of dependency treebanks across 47 languages with unified POS-tag and dependency label sets. Each treebank has designated training, development, and test portions. We experiment on Arabic, English, Chinese, Hindi, and Spanish (Table 2)—languages with diverse punctuation vocabularies and punctuation interaction rules, not to mention script directionality. For each treebank, we use the tokenization provided by UD, and take the punctuation tokens (which may be multi-character, such as …) to be the tokens with the PUNCT tag. We replace each straight double quotation mark " with either “ or ” as appropriate, and similarly for single quotation marks.¹² We split each non-punctuation token that ends in . (such as etc.) into a shorter non-punctuation token (etc) followed by a special punctuation token called the “abbreviation dot” (which is distinct from a period). We prepend a special punctuation mark ∘ to every sentence $x¯$⁠, which can serve to absorb an initial comma, for example.¹³ We then replace each token with the special symbol UNK if its type appeared fewer than 5 times in the training portion. This gives the surface sentences.

To estimate the vocabulary $V$ of underlying punctemes, we simply collect all surface token sequences x_i that appear at any slot in the training portion of the processed treebank. This is a generous estimate. Similarly, we estimate $Wd$ (§2.1) as all pairs $(l,r)∈V2$ that flank any d constituent.

Recall that our model generates surface punctuation given an unpunctuated dependency tree. We train it on each of the 5 languages independently. We evaluate on conditional perplexity, which will be low if the trained model successfully assigns a high probability to the actual surface punctuation in a held-out corpus of the same language.

We compare our model against three baselines to show that its complexity is necessary. Our first baseline is an ablation study that does not use latent underlying punctuation, but generates the surface punctuation directly from the tree. (To implement this, we fix the parameters of the noisy channel so that the surface punctuation equals the underlying with probability 1.) If our full model performs significantly better, it will demonstrate the importance of a distinct underlying layer.

Our other two baselines ignore the tree structure, so if our full model performs significantly better, it will demonstrate that conditioning on explicit syntactic structure is useful. These baselines are based on previously published approaches that reduce the problem to tagging: Xu et al. (2016) use a BiLSTM-CRF tagger with bigram topology; Tilk and Alumäe (2016) use a BiGRU tagger with attention. In both approaches, the model is trained to tag each slot i with the correct string $xi∈V*$ (possibly ϵ or ^∧). These are discriminative probabilistic models (in contrast to our generative one). Each gives a probability distribution over the taggings (conditioned on the unpunctuated sentence), so we can evaluate their perplexity.¹⁴

As shown in Table 3, our full model beats the baselines in perplexity in all 5 languages. Also, in 4 of 5 languages, allowing a trained NoisyChannel (rather than the identity map) significantly improves the perplexity.

We study our learned probability distribution over noisy channel rules (ab ↦ b, ab ↦ a, ab ↦ ab, ab ↦ ba) for English. The probability distributions corresponding to six of Nunberg’s English rules are shown in Figure 3. By comparing the orange and blue bars, observe that the model trained on the 𝖾𝗇_𝖼𝖾𝗌𝗅 treebank learned different quotation rules from the one trained on the 𝖾𝗇 treebank. This is because 𝖾𝗇_𝖼𝖾𝗌𝗅 follows British style, whereas 𝖾𝗇 has American-style quote transposition.¹⁵

We now focus on the model learned from the 𝖾𝗇 treebank. Nunberg’s rules are deterministic, and our noisy channel indeed learned low-entropy rules, in the sense that for an input ab with underlying count ≥ 25,¹⁶ at least one of the possible outputs (a, b, ab or ba) always has probability >0.75. The one exception is ”.↦.” for which the argmax output has probability ≈ 0.5, because writers do not apply this quote transposition rule consistently. As shown by the blue bars in Figure 3, the high-probability transduction rules are consistent with Nunberg’s hand-crafted deterministic grammar in Table 1.

Our system has high precision when we look at the confident rules. Of the 24 learned edits with conditional probability >0.75, Nunberg lists 20.

Our system also has good recall. Nunberg’s hand-crafted schemata consider 16 punctuation types and generate a total of 192 edit rules, including the specimens in Table 1. That is, of the 16² = 256 possible underlying punctuation bigrams ab, $34$ are supposed to undergo absorption or transposition. Our method achieves fairly high recall, in the sense that when Nunberg proposes ab ↦ γ, our learned p(γ∣ab) usually ranks highly among all probabilities of the form p(γ′∣ab). 75 of Nunberg’s rules got rank 1, 48 got rank 2, and the remaining 69 got rank >2. The mean reciprocal rank was 0.621. Recall is quite high when we restrict to those Nunberg rules ab ↦ γ for which our model is confident how to rewrite ab, in the sense that some p(γ′∣ab) > 0.5. (This tends to eliminate rare ab: see §5.) Of these 55 Nunberg rules, 38 rules got rank 1, 15 got rank 2, and only 2 got rank worse than 2. The mean reciprocal rank was 0.836.

¿What about Spanish? Spanish uses inverted question marks ¿ and exclamation marks ¡, which form symmetric pairs with the regular question marks and exclamation marks. If we try to extrapolate to Spanish from Nunberg’s English formalization, the English mark most analogous to ¿ is (. Our learned noisy channel for Spanish (not graphed here) includes the high-probability rules ,¿↦,¿ and :¿↦:¿ and ¿,↦¿ which match Nunberg’s treatment of ( in English.

What does our model learn about how dependency relations are marked by underlying punctuation?

The above example¹⁷ illustrates the use of specific puncteme pairs to set off the advmod, ccomp, and nmod relations. Notice that said takes a complement (ccomp) that is symmetrically quoted but also left delimited by a comma, which is indeed how direct speech is punctuated in English. This example also illustrates quotation transposition. The top five relations that are most likely to generate symmetric punctemes and their top (l, r) pairs are shown in Table 4.

The above example¹⁸ shows how our model handles commas in conjunctions of 2 or more phrases. UD format dictates that each conjunct after the first is attached by the conj relation. As shown above, each such conjunct is surrounded by underlying commas (via the N.,.,.conj feature from Appendix A), except for the one that bears the conjunction and (via an even stronger weight on the C.ε.ε.$conj→$⁠.cc feature). Our learned feature weights indeed yield p(ℓ = ε, r = ε) > 0.5 for the final conjunct in this example. Some writers omit the “Oxford comma” before the conjunction: this style can be achieved simply by changing “surrounded” to “preceded” (that is, changing the N feature to N.,.ε.conj).

We evaluate the trained punctuation model by using it in the following three tasks.

In this task, we are given a depunctuated sentence $d¯$¹⁹ and must restore its (surface) punctuation. Our model supposes that the observed punctuated sentence $x¯$ would have arisen via the generative process (1). Thus, we try to find T, T′, and $x¯$ that are consistent with $d¯$ (a partial observation of $x¯$⁠).

The first step is to reconstruct T from $d¯$⁠. This initial parsing step is intended to choose the T that maximizes $psyn(T∣d¯)$⁠.²⁰ This step depends only on p_syn and not on our punctuation model (p_θ, p_ϕ). In practice, we choose T via a dependency parser that has been trained on an unpunctuated treebank with examples of the form $(d¯,T)$⁠.²¹

We evaluate on Arabic, English, Chinese, Hindi, and Spanish. For each language, we train both the parser and the punctuation model on the training split of that UD treebank (§4), and evaluate on held-out data. We compare to the BiLSTM-CRF baseline in §4 (Xu et al., 2016).²² We also compare to a “trivial” deterministic baseline, which merely places a period at the end of the sentence (or a "|" in the case of Hindi) and adds no other punctuation. Because most slots do not in fact have punctuation, the trivial baseline already does very well; to improve on it, we must fix its errors without introducing new ones.

Our final comparison on test data is shown in the table in Figure 4. On all 5 languages, our method beats (usually significantly) its 3 competitors: the trivial deterministic baseline, the BiLSTM-CRF, and the ablated version of our model (Attach) that omits the noisy channel.

Of course, the success of our method depends on the quality of the parse trees T (which is particularly low for Chinese and Arabic). The graph in Figure 4 explores this relationship, by evaluating (on dev data) with noisier trees obtained from parsers that were variously trained on only the first 10%, 20%, …of the training data. On all 5 languages, provided that the trees are at least 75% correct, our punctuation model beats both the trivial baseline and the BiLSTM-CRF (which do not use trees). It also beats the Attach ablation baseline at all levels of tree accuracy (these curves are omitted from the graph to avoid clutter). In all languages, better parses give better performance, and gold trees yield the best results.

Our next goal is to correct punctuation errors in a learner corpus. Each sentence is drawn from the Cambridge Learner Corpus treebanks, which provide original (𝖾𝗇_𝖾𝗌𝗅) and corrected (𝖾𝗇_𝖼𝖾𝗌𝗅) sentences. All kinds of errors are corrected, such as syntax errors, but we use only the 30% of sentences whose depunctuated trees T are isomorphic between 𝖾𝗇_𝖾𝗌𝗅 and 𝖾𝗇_𝖼𝖾𝗌𝗅. These 𝖾𝗇_𝖼𝖾𝗌𝗅 trees may correct word and/or punctuation errors in 𝖾𝗇_𝖾𝗌𝗅, as we wish to do automatically.

We assume that an English learner can make mistakes in both the attachment and the noisy channel steps. A common attachment mistake is the failure to surround a non-restrictive relative clause with commas. In the noisy channel step, mistakes in quote transposition are common.

Based on the assumption about the two error sources, we develop a discriminative model for this task. Let $x¯e$ denote the full input sentence, and let x_e and x_c denote the input (possibly errorful) and output (corrected) punctuation sequences. We model $p(xc∣x¯e)=∑T∑Tc′psyn(T∣x¯e)⋅pθ(Tc′∣T,xe)⋅pϕ(xc∣Tc′)$⁠. Here T is the depunctuated parse tree, T_c′ is the corrected underlying tree, T_e′ is the error underlying tree, and we assume $pθ(Tc′∣T,xe)=∑Te′p(Te′∣T,xe)⋅pθ(Tc′∣Te′)$⁠.

Finally, we reconstruct x_c based on the noisy channel p_ϕ(x_c∣T_c′) in §2.2. During training, ϕ is regularized to be close to the noisy channel parameters in the punctuation model trained on 𝖾𝗇_𝖼𝖾𝗌𝗅.

We use the same MBR decoder as in §6.1 to choose the best action. We evaluate using AED as in §6.1. As a second metric, we use the script from the CoNLL 2014 Shared Task on Grammatical Error Correction (Ng et al., 2014): it computes the F_0.5-measure of the set of edits found by the system, relative to the true set of edits.

As shown in Table 5, our method achieves better performance than the punctuation restoration baselines (which ignore input punctuation). On the other hand, it is soundly beaten by a new BiLSTM-CRF that we trained specifically for the task of punctuation correction. This is the same as the BiLSTM-CRF in the previous section, except that the BiLSTM now reads a punctuated input sentence (with possibly erroneous punctuation). To be precise, at step 0 ≤ i ≤ n, the BiLSTM reads a concatenation of the embedding of word i (or BOS if i = 0) with an embedding of the punctuation token sequence x_i. The BiLSTM-CRF wins because it is a discriminative model tailored for this task: the BiLSTM can extract arbitrary contextual features of slot i that are correlated with whether x_i is correct in context.

We suspect that syntactic transformations on a sentence should often preserve the underlying punctuation attached to its tree. The surface punctuation can then be regenerated from the transformed tree. Such transformations include edits that are suggested by a writing assistance tool (Heidorn, 2000), or subtree deletions in compressive summarization (Knight and Marcu, 2002).

For our experiment, we evaluate an interesting case of syntactic transformation. Wang and Eisner (2016) consider a systematic rephrasing procedure by rearranging the order of dependent subtrees within a UD treebank, in order to synthesize new languages with different word order that can then be used to help train multi-lingual systems (i.e., data augmentation with synthetic data).

As Wang and Eisner acknowledge (2016, footnote 9), their permutations treat surface punctuation tokens like ordinary words, which can result in synthetic sentences whose punctuation is quite unlike that of real languages.

In our experiment, we use Wang and Eisner (2016) “self-permutation” setting, where the dependents of each noun and verb are stochastically reordered, but according to a dependent ordering model that has been trained on the same language. For example, rephrasing a English sentence

under an English ordering model may yield

which is still grammatical except that , and . are wrongly swapped (after all, they have the same POS tag and relation type). Worse, permutation may yield bizarre punctuation such as , , at the start of a sentence.

Our punctuation model gives a straightforward remedy—instead of permuting the tree directly, we first discover its most likely underlying tree

by the maximizing variant of Algorithm 1 (§3.1). Then, we permute the underlying tree and sample the surface punctuation from the distribution modeled by the trained PFST, yielding

We leave the handling of capitalization to future work.

We test the naturalness of the permuted sentences by asking how well a word trigram language model trained on them could predict the original sentences.²³ As shown in Figure 6, our permutation approach reduces the perplexity over the baseline on 4 of the 5 languages, often dramatically.

Punctuation can aid syntactic analysis, since it signals phrase boundaries and sentence structure. Briscoe (1994) and White and Rajkumar (2008) parse punctuated sentences using hand-crafted constraint-based grammars that implement Nunberg’s approach in a declarative way. These grammars treat surface punctuation symbols as ordinary words, but annotate the nonterminal categories so as to effectively keep track of the underlying punctuation. This is tantamount to crafting a grammar for underlyingly punctuated sentences and composing it with a finite-state noisy channel.

The parser of Ma et al. (2014) takes a different approach and treats punctuation marks as features of their neighboring words. Zhang et al. (2013) use a generative model for punctuated sentences, leting them restore punctuation marks during transition-based parsing of unpunctuated sentences. Li et al. (2005) use punctuation marks to segment a sentence: this “divide and rule” strategy reduces ambiguity in parsing of long Chinese sentences. Punctuation can similarly be used to constrain syntactic structure during grammar induction (Spitkovsky et al., 2011).

Punctuation restoration (§6.1) is useful for transcribing text from unpunctuated speech. The task is usually treated by tagging each slot with zero or more punctuation tokens, using a traditional sequence labeling method: conditional random fields (Lui and Wang, 2013; Lu and Ng, 2010), recurrent neural networks (Tilk and Alumäe, 2016), or transition-based systems (Ballesteros and Wanner, 2016).

We have provided a new computational approach to modeling punctuation. In our model, syntactic constituents stochastically generate latent underlying left and right punctemes. Surface punctuation marks are not directly attached to the syntax tree, but are generated from sequences of adjacent punctemes by a (stochastic) finite-state string rewriting process . Our model is inspired by Nunberg (1990) formal grammar for English punctuation, but is probabilistic and trainable. We give exact algorithms for training and inference.

We trained Nunberg-like models for 5 languages and L2 English. We compared the English model to Nunberg’s, and showed how the trained models can be used across languages for punctuation restoration, correction, and adjustment.

In the future, we would like to study the usefulness of the recovered underlying trees on tasks such as syntactically sensitive sentiment analysis (Tai et al., 2015), machine translation (Cowan et al., 2006), relation extraction (Culotta and Sorensen, 2004), and coreference resolution (Kong et al., 2010). We would also like to investigate how underlying punctuation could aid parsing. For discriminative parsing, features for scoring the tree could refer to the underlying punctuation, not just the surface punctuation. For generative parsing (§3), we could follow the scheme in (1). For example, the p_syn factor in (1) might be a standard recurrent neural network grammar (RNNG) (Dyer et al., 2016); when a subtree of T is completed by the Reduce operation of p_syn, the punctuation-augmented RNNG (1) would stochastically attach subtree-external left and right punctemes with p_θ and transduce the subtree-internal slots with p_ϕ.

In the future, we are also interested in enriching the T′ representation and making it more different from T, to underlyingly account for other phenomena in T such as capitalization, spacing, morphology, and non-projectivity (via reordering).

This material is based upon work supported by the National Science Foundation under Grant Nos. 1423276 and 1718846, including a REU supplement to the first author. We are grateful to the state of Maryland for the Maryland Advanced Research Computing Center, a crucial resource. We thank Xiaochen Li for early discussion, Argo lab members for further discussion, and the three reviewers for quality comments.