{ "paper_id": "P09-1012", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T08:53:18.733486Z" }, "title": "Bayesian Unsupervised Word Segmentation with Nested Pitman-Yor Language Modeling", "authors": [ { "first": "Daichi", "middle": [], "last": "Mochihashi", "suffix": "", "affiliation": {}, "email": "daichi@cslab.kecl.ntt.co.jp" }, { "first": "Takeshi", "middle": [], "last": "Yamada", "suffix": "", "affiliation": {}, "email": "yamada@cslab.kecl.ntt.co.jp" }, { "first": "Naonori", "middle": [], "last": "Ueda", "suffix": "", "affiliation": {}, "email": "ueda@cslab.kecl.ntt.co.jp" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "In this paper, we propose a new Bayesian model for fully unsupervised word segmentation and an efficient blocked Gibbs sampler combined with dynamic programming for inference. Our model is a nested hierarchical Pitman-Yor language model, where Pitman-Yor spelling model is embedded in the word model. We confirmed that it significantly outperforms previous reported results in both phonetic transcripts and standard datasets for Chinese and Japanese word segmentation. Our model is also considered as a way to construct an accurate word n-gram language model directly from characters of arbitrary language, without any \"word\" indications.", "pdf_parse": { "paper_id": "P09-1012", "_pdf_hash": "", "abstract": [ { "text": "In this paper, we propose a new Bayesian model for fully unsupervised word segmentation and an efficient blocked Gibbs sampler combined with dynamic programming for inference. Our model is a nested hierarchical Pitman-Yor language model, where Pitman-Yor spelling model is embedded in the word model. We confirmed that it significantly outperforms previous reported results in both phonetic transcripts and standard datasets for Chinese and Japanese word segmentation. Our model is also considered as a way to construct an accurate word n-gram language model directly from characters of arbitrary language, without any \"word\" indications.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "\"Word\" is no trivial concept in many languages. Asian languages such as Chinese and Japanese have no explicit word boundaries, thus word segmentation is a crucial first step when processing them. Even in western languages, valid \"words\" are often not identical to space-separated tokens. For example, proper nouns such as \"United Kingdom\" or idiomatic phrases such as \"with respect to\" actually function as a single word, and we often condense them into the virtual words \"UK\" and \"w.r.t.\".", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In order to extract \"words\" from text streams, unsupervised word segmentation is an important research area because the criteria for creating supervised training data could be arbitrary, and will be suboptimal for applications that rely on segmentations. It is particularly difficult to create \"correct\" training data for speech transcripts, colloquial texts, and classics where segmentations are often ambiguous, let alone is impossible for unknown languages whose properties computational linguists might seek to uncover.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "From a scientific point of view, it is also interesting because it can shed light on how children learn \"words\" without the explicitly given boundaries for every word, which is assumed by supervised learning approaches.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Lately, model-based methods have been introduced for unsupervised segmentation, in particular those based on Dirichlet processes on words (Goldwater et al., 2006; Xu et al., 2008) . This maximizes the probability of word segmentation w given a string s :", "cite_spans": [ { "start": 138, "end": 162, "text": "(Goldwater et al., 2006;", "ref_id": "BIBREF7" }, { "start": 163, "end": 179, "text": "Xu et al., 2008)", "ref_id": "BIBREF25" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "w = argmax w p(w|s) .", "eq_num": "(1)" } ], "section": "Introduction", "sec_num": "1" }, { "text": "This approach often implicitly includes heuristic criteria proposed so far 1 , while having a clear statistical semantics to find the most probable word segmentation that will maximize the probability of the data, here the strings. However, they are still na\u00efve with respect to word spellings, and the inference is very slow owing to inefficient Gibbs sampling. Crucially, since they rely on sampling a word boundary between two neighboring words, they can leverage only up to bigram word dependencies.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In this paper, we extend this work to propose a more efficient and accurate unsupervised word segmentation that will optimize the performance of the word n-gram Pitman-Yor (i.e. Bayesian Kneser-Ney) language model, with an accurate character \u221e-gram Pitman-Yor spelling model embedded in word models.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Furthermore, it can be viewed as a method for building a high-performance n-gram language model directly from character strings of arbitrary language. It is carefully smoothed and has no \"unknown words\" problem, resulting from its model structure.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "This paper is organized as follows. In Section 2, (a) Generating n-gram distributions G hierarchically from the Pitman-Yor process. Here, n = 3.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "(b) Equivalent representation using a hierarchical Chinese Restaurant process. Each word in a training text is a \"customer\" shown in italic, and added to the leaf of its two words context. we briefly describe a language model based on the Pitman-Yor process (Teh, 2006b) , which is a generalization of the Dirichlet process used in previous research. By embedding a character n-gram in word n-gram from a Bayesian perspective, Section 3 introduces a novel language model for word segmentation, which we call the Nested Pitman-Yor language model. Section 4 describes an efficient blocked Gibbs sampler that leverages dynamic programming for inference. In Section 5 we describe experiments on the standard datasets in Chinese and Japanese in addition to English phonetic transcripts, and semi-supervised experiments are also explored. Section 6 is a discussion and Section 7 concludes the paper.", "cite_spans": [ { "start": 258, "end": 270, "text": "(Teh, 2006b)", "ref_id": "BIBREF23" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "To compute a probability p(w|s) in (1), we adopt a Bayesian language model lately proposed by (Teh, 2006b; Goldwater et al., 2005 ) based on the Pitman-Yor process, a generalization of the Dirichlet process. As we shall see, this is a Bayesian theory of the best-performing Kneser-Ney smoothing of n-grams (Kneser and Ney, 1995) , allowing an integrated modeling from a Bayesian perspective as persued in this paper. The Pitman-Yor (PY) process is a stochastic process that generates discrete probability distribution G that is similar to another distribution G 0 , called a base measure. It is written as", "cite_spans": [ { "start": 94, "end": 106, "text": "(Teh, 2006b;", "ref_id": "BIBREF23" }, { "start": 107, "end": 129, "text": "Goldwater et al., 2005", "ref_id": "BIBREF6" }, { "start": 306, "end": 328, "text": "(Kneser and Ney, 1995)", "ref_id": "BIBREF13" } ], "ref_spans": [], "eq_spans": [], "section": "Pitman-Yor process and n-gram models", "sec_num": "2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "G \u223c PY(G 0 , d, \u03b8) ,", "eq_num": "(2)" } ], "section": "Pitman-Yor process and n-gram models", "sec_num": "2" }, { "text": "where d is a discount factor and \u03b8 controls how similar G is to G 0 on average. Suppose we have a unigram word distribution", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Pitman-Yor process and n-gram models", "sec_num": "2" }, { "text": "G 1 = { p(\u2022) } where \u2022 ranges over each word in the lexicon. The bigram distribution G 2 = { p(\u2022|v) }", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Pitman-Yor process and n-gram models", "sec_num": "2" }, { "text": "given a word v is different from G 1 , but will be similar to G 1 especially for high frequency words. Therefore, we can generate G 2 from a PY process of base measure G 1 , as", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Pitman-Yor process and n-gram models", "sec_num": "2" }, { "text": "G 2 \u223c PY(G 1 , d, \u03b8). Similarly, trigram distribution G 3 = { p(\u2022|v v) } given an additional word v is generated as G 3 \u223c PY(G 2 , d, \u03b8)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Pitman-Yor process and n-gram models", "sec_num": "2" }, { "text": ", and G 1 , G 2 , G 3 will form a tree structure shown in Figure 1(a) . In practice, we cannot observe G directly because it will be infinite dimensional distribution over the possible words, as we shall see in this paper. However, when we integrate out G it is known that Figure 1 (a) can be represented by an equivalent hierarchical Chinese Restaurant Process (CRP) (Aldous, 1985) as in Figure 1(b) .", "cite_spans": [ { "start": 368, "end": 382, "text": "(Aldous, 1985)", "ref_id": "BIBREF0" } ], "ref_spans": [ { "start": 58, "end": 69, "text": "Figure 1(a)", "ref_id": "FIGREF0" }, { "start": 273, "end": 281, "text": "Figure 1", "ref_id": "FIGREF0" }, { "start": 389, "end": 400, "text": "Figure 1(b)", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "Pitman-Yor process and n-gram models", "sec_num": "2" }, { "text": "In this representation, each n-gram context h (including the null context for unigrams) is a Chinese restaurant whose customers are the n-gram counts c(w|h) seated over the tables", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Pitman-Yor process and n-gram models", "sec_num": "2" }, { "text": "1 \u2022 \u2022 \u2022 t hw .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Pitman-Yor process and n-gram models", "sec_num": "2" }, { "text": "The seatings has been incrementally constructed by choosing the table k for each count in c(w|h) with probability proportional to", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Pitman-Yor process and n-gram models", "sec_num": "2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "c hwk \u2212 d (k = 1, \u2022 \u2022 \u2022 , t hw ) \u03b8 + d\u2022t h\u2022 (k = new ) ,", "eq_num": "(3)" } ], "section": "Pitman-Yor process and n-gram models", "sec_num": "2" }, { "text": "where c hwk is the number of customers seated at table k thus far and t h\u2022 = w t hw is the total number of tables in h. When k = new is selected, t hw is incremented, and this means that the count was actually generated from the shorter context h . Therefore, in that case a proxy customer is sent to the parent restaurant and this process will recurse. For example, if we have a sentence \"she will sing\" in the training data for trigrams, we add each word \"she\" \"will\" \"sing\" \"$\" as a customer to its two preceding words context node, as described in Figure 1 (b). Here, \"$\" is a special token representing a sentence boundary in language model-ing (Brown et al., 1992) .", "cite_spans": [ { "start": 650, "end": 670, "text": "(Brown et al., 1992)", "ref_id": "BIBREF2" } ], "ref_spans": [ { "start": 552, "end": 560, "text": "Figure 1", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "Pitman-Yor process and n-gram models", "sec_num": "2" }, { "text": "As a result, the n-gram probability of this hierarchical Pitman-Yor language model (HPYLM) is recursively computed as", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Pitman-Yor process and n-gram models", "sec_num": "2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "p(w|h) = c(w|h)\u2212d\u2022t hw \u03b8+c(h) + \u03b8+d\u2022t h\u2022 \u03b8+c(h) p(w|h ),", "eq_num": "(4)" } ], "section": "Pitman-Yor process and n-gram models", "sec_num": "2" }, { "text": "where p(w|h ) is the same probability using a (n\u22121)-gram context h . When we set t hw \u2261 1, (4) recovers a Kneser-Ney smoothing: thus a HPYLM is a Bayesian Kneser-Ney language model as well as an extension of the hierarchical Dirichlet Process (HDP) used in Goldwater et al. (2006) . \u03b8, d are hyperparameters that can be learned as Gamma and Beta posteriors, respectively, given the data. For details, see Teh (2006a) .", "cite_spans": [ { "start": 257, "end": 280, "text": "Goldwater et al. (2006)", "ref_id": "BIBREF7" }, { "start": 405, "end": 416, "text": "Teh (2006a)", "ref_id": "BIBREF22" } ], "ref_spans": [], "eq_spans": [], "section": "Pitman-Yor process and n-gram models", "sec_num": "2" }, { "text": "The inference of this model interleaves adding and removing a customer to optimize t hw , d, and \u03b8 using MCMC. However, in our case \"words\" are not known a priori: the next section describes how to accomplish this by constructing a nested HPYLM of words and characters, with the associated inference algorithm.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Pitman-Yor process and n-gram models", "sec_num": "2" }, { "text": "Thus far we have assumed that the unigram G 1 is already given, but of course it should also be generated as", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Nested Pitman-Yor Language Model", "sec_num": "3" }, { "text": "G 1 \u223c PY(G 0 , d, \u03b8).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Nested Pitman-Yor Language Model", "sec_num": "3" }, { "text": "Here, a problem occurs: What should we use for G 0 , namely the prior probabilities over words 2 ? If a lexicon is finite, we can use a uniform prior G 0 (w) = 1/|V | for every word w in lexicon V . However, with word segmentation every substring could be a word, thus the lexicon is not limited but will be countably infinite.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Nested Pitman-Yor Language Model", "sec_num": "3" }, { "text": "Building an accurate G 0 is crucial for word segmentation, since it determines how the possible words will look like. Previous work using a Dirichlet process used a relatively simple prior for G 0 , namely an uniform distribution over characters (Goldwater et al., 2006) , or a prior solely dependent on word length with a Poisson distribution whose parameter is fixed by hand (Xu et al., 2008) .", "cite_spans": [ { "start": 246, "end": 270, "text": "(Goldwater et al., 2006)", "ref_id": "BIBREF7" }, { "start": 377, "end": 394, "text": "(Xu et al., 2008)", "ref_id": "BIBREF25" } ], "ref_spans": [], "eq_spans": [], "section": "Nested Pitman-Yor Language Model", "sec_num": "3" }, { "text": "In contrast, in this paper we use a simple but more elaborate model, that is, a character n-gram language model that also employs HPYLM. This is important because in English, for example, words are likely to end in '-tion' and begin with 're-', but almost never end in '-tio' nor begin with 'sre-' 3 .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Nested Pitman-Yor Language Model", "sec_num": "3" }, { "text": "Therefore, we use", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Nested Pitman-Yor Language Model", "sec_num": "3" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "G 0 (w) = p(c 1 \u2022 \u2022 \u2022 c k ) (5) = k i=1 p(c i |c 1 \u2022 \u2022 \u2022 c i\u22121 )", "eq_num": "(6)" } ], "section": "Nested Pitman-Yor Language Model", "sec_num": "3" }, { "text": "where string c 1 \u2022 \u2022 \u2022 c k is a spelling of w, and", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Nested Pitman-Yor Language Model", "sec_num": "3" }, { "text": "p(c i |c 1 \u2022 \u2022 \u2022 c i\u22121 )", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Nested Pitman-Yor Language Model", "sec_num": "3" }, { "text": "is given by the character HPYLM according to (4). This language model, which we call Nested Pitman-Yor Language Model (NPYLM) hereafter, is the hierarchical language model shown in Figure 2, where the character HPYLM is embedded as a base measure of the word HPYLM. 4 As the final base measure for the character HPYLM, we used a uniform prior over the possible characters of a given language. To avoid dependency on ngram order n, we actually used the \u221e-gram language model (Mochihashi and Sumita, 2007) , a variable order HPYLM, for characters. However, for generality we hereafter state that we used the HPYLM. The theory remains the same for \u221egrams, except sampling or marginalizing over n as needed.", "cite_spans": [ { "start": 474, "end": 503, "text": "(Mochihashi and Sumita, 2007)", "ref_id": "BIBREF16" } ], "ref_spans": [ { "start": 181, "end": 187, "text": "Figure", "ref_id": null } ], "eq_spans": [], "section": "Nested Pitman-Yor Language Model", "sec_num": "3" }, { "text": "Furthermore, we corrected (5) so that word length will have a Poisson distribution whose parameter can now be estimated for a given language and word type. We describe this in detail in Section 4.3.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Nested Pitman-Yor Language Model", "sec_num": "3" }, { "text": "In our NPYLM, the word model and the character model are not separate but connected through a nested CRP. When a word w is generated from its parent at the unigram node, it means that w is drawn from the base measure, namely a character HPYLM. Then we divide w into characters c 1 \u2022 \u2022 \u2022 c k to yield a \"sentence\" of characters and feed this into the character HPYLM as data.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Chinese Restaurant Representation", "sec_num": null }, { "text": "Conversely, when a table becomes empty, this means that the data associated with the table are no longer valid. Therefore we remove the corresponding customers from the character HPYLM using the inverse procedure of adding a customer in Section 2.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Chinese Restaurant Representation", "sec_num": null }, { "text": "All these processes will be invoked when a string is segmented into \"words\" and customers are added to the leaves of the word HPYLM. To segment a string into \"words\", we used efficient dynamic programming combined with MCMC, as described in the next section.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Chinese Restaurant Representation", "sec_num": null }, { "text": "To find the hidden word segmentation w of a string", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Inference", "sec_num": "4" }, { "text": "s = c 1 \u2022 \u2022 \u2022 c N ,", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Inference", "sec_num": "4" }, { "text": "which is equivalent to the vector of binary hidden variables z = z 1 \u2022 \u2022 \u2022 z N , the simplest approach is to build a Gibbs sampler that randomly selects a character c i and draw a binary decision z i as to whether there is a word boundary, and then update the language model according to the new segmentation (Goldwater et al., 2006; Xu et al., 2008) . When we iterate this procedure sufficiently long, it becomes a sample from the true distribution (1) (Gilks et al., 1996) .", "cite_spans": [ { "start": 309, "end": 333, "text": "(Goldwater et al., 2006;", "ref_id": "BIBREF7" }, { "start": 334, "end": 350, "text": "Xu et al., 2008)", "ref_id": "BIBREF25" }, { "start": 454, "end": 474, "text": "(Gilks et al., 1996)", "ref_id": "BIBREF5" } ], "ref_spans": [], "eq_spans": [], "section": "Inference", "sec_num": "4" }, { "text": "However, this sampler is too inefficient since time series data such as word segmentation have a very high correlation between neighboring words. As a result, the sampler is extremely slow to converge. In fact, (Goldwater et al., 2006) reports that the sampler would not mix without annealing, and the experiments needed 20,000 times of sampling for every character in the training data.", "cite_spans": [ { "start": 211, "end": 235, "text": "(Goldwater et al., 2006)", "ref_id": "BIBREF7" } ], "ref_spans": [], "eq_spans": [], "section": "Inference", "sec_num": "4" }, { "text": "Furthermore, it has an inherent limitation that it cannot deal with larger than bigrams, because it uses only local statistics between directly contiguous words for word segmentation.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Inference", "sec_num": "4" }, { "text": "Instead, we propose a sentence-wise Gibbs sampler of word segmentation using efficient dynamic programming, as shown in Figure 3 .", "cite_spans": [], "ref_spans": [ { "start": 120, "end": 128, "text": "Figure 3", "ref_id": null } ], "eq_spans": [], "section": "Blocked Gibbs sampler", "sec_num": "4.1" }, { "text": "In this algorithm, first we randomly select a string, and then remove the \"sentence\" data of its word segmentation from the NPYLM. Sampling a new segmentation, we update the NPYLM by adding a new \"sentence\" according to the new seg-", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Blocked Gibbs sampler", "sec_num": "4.1" }, { "text": "1: for j = 1 \u2022 \u2022 \u2022 J do 2: for s in randperm (s 1 , \u2022 \u2022 \u2022 , s D ) do 3: if j > 1 then 4:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Blocked Gibbs sampler", "sec_num": "4.1" }, { "text": "Remove customers of w(s) from \u0398 Sample hyperparameters of \u0398 10: end for Figure 3 : Blocked Gibbs Sampler of NPYLM \u0398.", "cite_spans": [], "ref_spans": [ { "start": 72, "end": 80, "text": "Figure 3", "ref_id": null } ], "eq_spans": [], "section": "Blocked Gibbs sampler", "sec_num": "4.1" }, { "text": "mentation. When we repeat this process, it is expected to mix rapidly because it implicitly considers all possible segmentations of the given string at the same time. This is called a blocked Gibbs sampler that samples z block-wise for each sentence. It has an additional advantage in that we can accommodate higher-order relationships than bigrams, particularly trigrams, for word segmentation. 5", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Blocked Gibbs sampler", "sec_num": "4.1" }, { "text": "Then, how can we sample a segmentation w for each string s? In accordance with the Forward filtering Backward sampling of HMM (Scott, 2002) , this is achieved by essentially the same algorithm employed to sample a PCFG parse tree within MCMC (Johnson et al., 2007) and grammar-based segmentation . Figure 4) . Segmentations before the final k characters are marginalized using the following recursive relationship:", "cite_spans": [ { "start": 126, "end": 139, "text": "(Scott, 2002)", "ref_id": "BIBREF20" }, { "start": 242, "end": 264, "text": "(Johnson et al., 2007)", "ref_id": "BIBREF12" } ], "ref_spans": [ { "start": 298, "end": 307, "text": "Figure 4)", "ref_id": "FIGREF5" } ], "eq_spans": [], "section": "Forward-Backward inference", "sec_num": "4.2" }, { "text": "\u03b1[t][k] = t\u2212k j=1 p(c t t\u2212k+1 |c t\u2212k t\u2212k\u2212j+1 )\u2022\u03b1[t\u2212k][j] (7)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forward-Backward inference", "sec_num": "4.2" }, { "text": "where ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forward-Backward inference", "sec_num": "4.2" }, { "text": "for k = max(1, t\u2212L) to t do 3:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forward-Backward inference", "sec_num": "4.2" }, { "text": "Compute \u03b1[t][k] according to (7).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forward-Backward inference", "sec_num": "4.2" }, { "text": "end for 5: end for 6: Initialize t \u2190 N , i \u2190 0, w 0 \u2190 $ 7: while t > 0 do 8:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "4:", "sec_num": null }, { "text": "Draw k \u221d p(w i |c t t\u2212k+1 , \u0398) \u2022 \u03b1[t][k] 9: Set w i \u2190 c t t\u2212k+1", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "4:", "sec_num": null }, { "text": "10:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "4:", "sec_num": null }, { "text": "Set t \u2190 t \u2212 k, i \u2190 i + 1 11: end while 12: word boundary for each t as q t , we can write", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "4:", "sec_num": null }, { "text": "Return w = w i , w i\u22121 , \u2022 \u2022 \u2022 , w 1 .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "4:", "sec_num": null }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "\u03b1[t][k] = p(c t 1 , q t = k) (8) = j p(c t 1 , q t = k, q t\u2212k = j) (9) = j p(c t\u2212k 1 , c t t\u2212k+1 , q t = k, q t\u2212k = j)(10) = j p(c t t\u2212k+1 |c t\u2212k 1 )p(c t\u2212k 1 , q t\u2212k = j)(11) = j p(c t t\u2212k+1 |c t\u2212k )\u03b1[t\u2212k][j] ,", "eq_num": "(12)" } ], "section": "4:", "sec_num": null }, { "text": "where we used conditional independency of q t given q t\u2212k and uniform prior over q t in (11) above.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "4:", "sec_num": null }, { "text": "Backward Sampling. Once the probability table \u03b1[t][k] is obtained, we can sample a word segmentation backwards. Since \u03b1[N ][k] is a marginal probability of string c N 1 with the last k characters being a word, and there is always a sentence boundary token $ at the end of the string, with probability proportional to p", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "4:", "sec_num": null }, { "text": "($|c N N \u2212k )\u2022\u03b1[N ][k]", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "4:", "sec_num": null }, { "text": "we can sample k to choose the boundary of the final word. The second final word is similarly sampled using the probability of preceding the last word just sampled: we continue this process until we arrive at the beginning of the string (Figure 5 ).", "cite_spans": [], "ref_spans": [ { "start": 236, "end": 245, "text": "(Figure 5", "ref_id": "FIGREF6" } ], "eq_spans": [], "section": "4:", "sec_num": null }, { "text": "Trigram case. For simplicity, we showed the algorithm for bigrams above. For trigrams, we maintain a forward variable \u03b1[t][k][j], which represents a marginal probability of string c 1 \u2022 \u2022 \u2022 c t with both the final k characters and further j characters preceding it being words. Forward-Backward algorithm becomes complicated thus omitted, but can be derived following the extended algorithm for second order HMM (He, 1988) .", "cite_spans": [ { "start": 412, "end": 422, "text": "(He, 1988)", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "4:", "sec_num": null }, { "text": "Complexity This algorithm has a complexity of O(N L 2 ) for bigrams and O(N L 3 ) for trigrams for each sentence, where N is the length of the sentence and L is the maximum allowed length of a word (\u2264 N ).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "4:", "sec_num": null }, { "text": "As Nagata (1996) noted, when only (5) is used inadequately low probabilities are assigned to long words, because it has a largely exponential distribution over length. To correct this, we assume that word length k has a Poisson distribution with a mean \u03bb:", "cite_spans": [ { "start": 3, "end": 16, "text": "Nagata (1996)", "ref_id": "BIBREF18" } ], "ref_spans": [], "eq_spans": [], "section": "Poisson correction", "sec_num": "4.3" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "Po(k|\u03bb) = e \u2212\u03bb \u03bb k k! .", "eq_num": "(13)" } ], "section": "Poisson correction", "sec_num": "4.3" }, { "text": "Since the appearance of c 1 \u2022 \u2022 \u2022 c k is equivalent to that of length k and the content, by making the character n-gram model explicit as \u0398 we can set", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Poisson correction", "sec_num": "4.3" }, { "text": "p(c 1 \u2022 \u2022 \u2022 c k ) = p(c 1 \u2022 \u2022 \u2022 c k , k) (14) = p(c 1 \u2022 \u2022 \u2022 c k , k|\u0398) p(k|\u0398) Po(k|\u03bb) (15)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Poisson correction", "sec_num": "4.3" }, { "text": "where", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Poisson correction", "sec_num": "4.3" }, { "text": "p(c 1 \u2022 \u2022 \u2022 c k , k|\u0398)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Poisson correction", "sec_num": "4.3" }, { "text": "is an n-gram probability given by (6), and p(k|\u0398) is a probability that a word of length k will be generated from \u0398. While previous work used p(k|\u0398) = (1 \u2212 p($)) k\u22121 p($), this is only true for unigrams. Instead, we employed a Monte Carlo method that generates words randomly from \u0398 to obtain the empirical estimates of p(k|\u0398).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Poisson correction", "sec_num": "4.3" }, { "text": "Estimating \u03bb. Of course, we do not leave \u03bb as a constant. Instead, we put a Gamma distribution", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Poisson correction", "sec_num": "4.3" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "p(\u03bb) = Ga(a, b) = b a \u0393(a) \u03bb a\u22121 e \u2212b\u03bb", "eq_num": "(16)" } ], "section": "Poisson correction", "sec_num": "4.3" }, { "text": "to estimate \u03bb from the data for given language and word type. 7 Here, \u0393(x) is a Gamma function and a, b are the hyperparameters chosen to give a nearly uniform prior distribution. 8", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Poisson correction", "sec_num": "4.3" }, { "text": "Denoting W as a set of \"words\" obtained from word segmentation, the posterior distribution of \u03bb used for (13) is", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Poisson correction", "sec_num": "4.3" }, { "text": "p(\u03bb|W ) \u221d p(W |\u03bb)p(\u03bb) = Ga a+ w\u2208W t(w)|w|, b+ w\u2208W t(w) , (17)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Poisson correction", "sec_num": "4.3" }, { "text": "where t(w) is the number of times word w is generated from the character HPYLM, i.e. the number of tables t w for w in word unigrams. We sampled \u03bb from this posterior for each Gibbs iteration.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Poisson correction", "sec_num": "4.3" }, { "text": "To validate our model, we conducted experiments on standard datasets for Chinese and Japanese word segmentation that are publicly available, as well as the same dataset used in (Goldwater et al., 2006) . Note that NPYLM maximizes the probability of strings, equivalently, minimizes the perplexity per character. Therefore, the recovery of the \"ground truth\" that is not available for inference is a byproduct in unsupervised learning.", "cite_spans": [ { "start": 177, "end": 201, "text": "(Goldwater et al., 2006)", "ref_id": "BIBREF7" } ], "ref_spans": [], "eq_spans": [], "section": "Experiments", "sec_num": "5" }, { "text": "Since our implementation is based on Unicode and learns all hyperparameters from the data, we also confirmed that NPYLM segments the Arabic Gigawords equally well.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experiments", "sec_num": "5" }, { "text": "In order to directly compare with the previously reported result, we first used the same dataset as Goldwater et al. (2006) . This dataset consists of 9,790 English phonetic transcripts from CHILDES data (MacWhinney and Snow, 1985) .", "cite_spans": [ { "start": 100, "end": 123, "text": "Goldwater et al. (2006)", "ref_id": "BIBREF7" }, { "start": 204, "end": 231, "text": "(MacWhinney and Snow, 1985)", "ref_id": "BIBREF15" } ], "ref_spans": [], "eq_spans": [], "section": "English phonetic transcripts", "sec_num": "5.1" }, { "text": "Since our algorithm converges rather fast, we ran the Gibbs sampler of trigram NPYLM for 200 iterations to obtain the results in Table 1 . Among the token precision (P), recall (R), and F-measure (F), the recall is especially higher to outperform the previous result based on HDP in F-measure. Meanwhile, the same measures over the obtained lexicon (LP, LR, LF) are not always improved. Moreover, the average length of words inferred was surprisingly similar to ground truth: 2.88, while the ground truth is 2.87. Table 2 shows the empirical computational time needed to obtain these results. Although the convergence in MCMC is not uniquely identified, improvement in efficiency is also outstanding.", "cite_spans": [], "ref_spans": [ { "start": 129, "end": 136, "text": "Table 1", "ref_id": null }, { "start": 514, "end": 521, "text": "Table 2", "ref_id": null } ], "eq_spans": [], "section": "English phonetic transcripts", "sec_num": "5.1" }, { "text": "To show applicability beyond small phonetic transcripts, we used standard datasets for Chinese and Table 1 : Segmentation accuracies on English phonetic transcripts. NPY(n) means n-gram NPYLM. Results for HDP(2) are taken from , which corrects the errors in Goldwater et al. (2006) .", "cite_spans": [ { "start": 258, "end": 281, "text": "Goldwater et al. (2006)", "ref_id": "BIBREF7" } ], "ref_spans": [ { "start": 99, "end": 106, "text": "Table 1", "ref_id": null } ], "eq_spans": [], "section": "Chinese and Japanese word segmentation", "sec_num": "5.2" }, { "text": "time iterations NPYLM 17min 200 HDP 10h 55min 20000 Table 2 : Computations needed for Table 1 . Iterations for \"HDP\" is the same as described in . Actually, NPYLM approximately converged around 50 iterations, 4 minutes.", "cite_spans": [], "ref_spans": [ { "start": 52, "end": 59, "text": "Table 2", "ref_id": null }, { "start": 86, "end": 93, "text": "Table 1", "ref_id": null } ], "eq_spans": [], "section": "Model", "sec_num": null }, { "text": "Japanese word segmentation, with all supervised segmentations removed in advance.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Model", "sec_num": null }, { "text": "Chinese For Chinese, we used a publicly available SIGHAN Bakeoff 2005 dataset (Emerson, 2005 . To compare with the latest unsupervised results (using a closed dataset of Bakeoff 2006), we chose the common sets prepared by Microsoft Research Asia (MSR) for simplified Chinese, and by City University of Hong Kong (CITYU) for traditional Chinese. We used a random subset of 50,000 sentences from each dataset for training, and the evaluation was conducted on the enclosed test data. 9", "cite_spans": [ { "start": 50, "end": 69, "text": "SIGHAN Bakeoff 2005", "ref_id": null }, { "start": 70, "end": 92, "text": "dataset (Emerson, 2005", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Model", "sec_num": null }, { "text": "Japanese For Japanese, we used the Kyoto Corpus (Kyoto) (Kurohashi and Nagao, 1998) : we used random subset of 1,000 sentences for evaluation and the remaining 37,400 sentences for training. In all cases we removed all whitespaces to yield raw character strings for inference, and set L = 4 for Chinese and L = 8 for Japanese to run the Gibbs sampler for 400 iterations. The results (in token F-measures) are shown in Table 3 . Our NPYLM significantly ourperforms the best results using a heuristic approach reported in Zhao and Kit (2008) . While Japanese accuracies appear lower, subjective qualities are much higher. This is mostly because NPYLM segments inflectional suffixes and combines frequent proper names, which are inconsistent with the \"correct\" Model MSR CITYU Kyoto NPY(2) 80.2 (51.9) 82.4 (126.5) 62.1 (23.1) NPY(3) 80.7 (48.8) 81.7 (128.3) 66.6 (20.6) ZK08 66.7 (-) 69.2 (-) - Table 3 : Accuracies and perplexities per character (in parentheses) on actual corpora. \"ZK08\" are the best results reported in Zhao and Kit (2008) . We used \u221e-gram for characters. segmentations. Bigram and trigram performances are similar for Chinese, but trigram performs better for Japanese. In fact, although the difference in perplexity per character is not so large, the perplexity per word is radically reduced: 439.8 (bigram) to 190.1 (trigram). This is because trigram models can leverage complex dependencies over words to yield shorter words, resulting in better predictions and increased tokens.", "cite_spans": [ { "start": 56, "end": 83, "text": "(Kurohashi and Nagao, 1998)", "ref_id": "BIBREF14" }, { "start": 520, "end": 539, "text": "Zhao and Kit (2008)", "ref_id": "BIBREF26" }, { "start": 1021, "end": 1040, "text": "Zhao and Kit (2008)", "ref_id": "BIBREF26" } ], "ref_spans": [ { "start": 418, "end": 425, "text": "Table 3", "ref_id": null }, { "start": 893, "end": 900, "text": "Table 3", "ref_id": null } ], "eq_spans": [], "section": "Model", "sec_num": null }, { "text": "Furthermore, NPYLM is easily amenable to semi-supervised or even supervised learning. In that case, we have only to replace the word segmentation w(s) in Figure 3 to the supervised one, for all or part of the training data. Table 4 shows the results using 10,000 sentences (1/5) or complete supervision. Our completely generative model achieves the performance of 94% (Chinese) or even 97% (Japanese) in supervised case. The result also shows that the supervised segmentations are suboptimal with respect to the perplexity per character, and even worse than unsupervised results. In semi-supervised case, using only 10K reference segmentations gives a performance of around 90% accuracy and the lowest perplexity, thanks to a combination with unsupervised data in a principled fashion.", "cite_spans": [], "ref_spans": [ { "start": 154, "end": 162, "text": "Figure 3", "ref_id": null }, { "start": 224, "end": 231, "text": "Table 4", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "MSR", "sec_num": null }, { "text": "Our model is particularly effective for spoken transcripts, colloquial texts, classics, or unknown languages where supervised segmentation data is difficult or even impossible to create. For example, we are pleased to say that we can now analyze (and build a language model on) \"The Tale of Genji\", the core of Japanese classics written 1,000 years ago ( Figure 6 ). The inferred segmentations are Figure 6 : Unsupervised segmentation result for \"The Tale of Genji\". (16, 443 sentences, 899, 668 characters in total) mostly correct, with some inflectional suffixes being recognized as words, which is also the case with English. Finally, we note that our model is also effective for western languages: Figure 7 shows a training text of \"Alice in Wonderland \" with all whitespaces removed, and the segmentation result.", "cite_spans": [ { "start": 459, "end": 471, "text": "Genji\". (16,", "ref_id": null }, { "start": 472, "end": 486, "text": "443 sentences,", "ref_id": null }, { "start": 487, "end": 491, "text": "899,", "ref_id": null }, { "start": 492, "end": 516, "text": "668 characters in total)", "ref_id": null } ], "ref_spans": [ { "start": 355, "end": 363, "text": "Figure 6", "ref_id": null }, { "start": 398, "end": 406, "text": "Figure 6", "ref_id": null }, { "start": 702, "end": 710, "text": "Figure 7", "ref_id": "FIGREF9" } ], "eq_spans": [], "section": "Classics and English text", "sec_num": "5.3" }, { "text": "\u00a2 \u00a1 \u00a4 \u00a3 \u00a6 \u00a5 \u00a7 \u00a6 \u00a9 \u00a2 \u00a7 \u00a4 ! # \" % $ ' & ) ( 1 0 3 2 $ # \" % 2 4 # 5 7 6 \u00a2 8 ' @ 9 B A D C F E G 9 3 H P I Q S R 7 0 @ T V U W 8 X W Y \u00a3 b a V \u00a9 \u00a4 c d I W $ b \" f e g @ h i 4 i h D p % R F q c P r s h # t u R v 9 P w F 2 7 % U x h y $ # \" ) ' 5 \u00a7 7 i @ c \" G \u00cf 7 \u00a5 \u00a4 @ 9 \u00a2 3 c g f b $ b \" f e V p \u00a4 s q v \u00a3 ! r s h i F \u00a5 W b v $ # d 7 R 8 e \" F f a 7 A b X 1 0 @ g p i h F j S \u00a5 k f l u % m d 4 a i y \u2022 \u2022 \u2022", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Classics and English text", "sec_num": "5.3" }, { "text": "While the data is extremely small (only 1,431 lines, 115,961 characters), our trigram NPYLM can infer the words surprisingly well. This is because our model contains both word and character models that are combined and carefully smoothed, from a Bayesian perspective.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Classics and English text", "sec_num": "5.3" }, { "text": "In retrospect, our NPYLM is essentially a hierarchical Markov model where the units (=words) evolve as the Markov process, and each unit has subunits (=characters) that also evolve as the Markov process. Therefore, for such languages as English that have already space-separated tokens, we can also begin with tokens besides the character-based approach in Section 5.3. In this case, each token is a \"character\" whose code is the integer token type, and a sentence is a sequence of \"characters.\" Figure 8 shows a part of the result computed over 100K sentences from Penn Treebank. We can see that some frequent phrases are identified as \"words\", using a fully unsupervised approach. Notice that this is only attainable with NPYLM where each phrase is described as a ngram model on its own, here a word \u221e-gram language model. While we developed an efficient forwardbackward algorithm for unsupervised segmentation, it is reminiscent of CRF in the discriminative approach. Therefore, it is also interesting to combine them in a discriminative way as persued in POS tagging using CRF+HMM (Suzuki et al., 2007) , let alone a simple semi-supervised approach in Section 5.2. This paper provides a foundation of such possibilities.", "cite_spans": [ { "start": 1085, "end": 1106, "text": "(Suzuki et al., 2007)", "ref_id": "BIBREF21" } ], "ref_spans": [ { "start": 496, "end": 504, "text": "Figure 8", "ref_id": null } ], "eq_spans": [], "section": "Discussion", "sec_num": "6" }, { "text": "lastly,shepicturedtoherselfhowthissamelittlesisterofhersw ould,intheafter-time,beherselfagrownwoman;andhowshe wouldkeep,throughallherriperyears,thesimpleandlovingh eartofherchildhood:andhowshewouldgatheraboutherothe rlittlechildren,andmaketheireyesbrightandeagerwithmany astrangetale,perhapsevenwiththedreamofwonderlandoflo ngago:andhowshewouldfeelwithalltheirsimplesorrows,an dfindapleasureinalltheirsimplejoys,rememberingherownc hild-life,andthehappysummerdays.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion", "sec_num": "6" }, { "text": "(a) Training data (in part). last ly , she pictured to herself how this same little sister of her s would , inthe after -time , be herself agrown woman ; and how she would keep , through allher ripery ears , the simple and loving heart of her child hood : and how she would gather about her other little children ,and make theireyes bright and eager with many a strange tale , perhaps even with the dream of wonderland of longago : and how she would feel with all their simple sorrow s , and find a pleasure in all their simple joys , remember ing her own child -life , and thehappy summerday s . ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion", "sec_num": "6" }, { "text": "In this paper, we proposed a much more efficient and accurate model for fully unsupervised word segmentation. With a combination of dynamic programming and an accurate spelling model from a Bayesian perspective, our model significantly outperforms the previous reported results, and the inference is very efficient.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "7" }, { "text": "This model is also considered as a way to build a Bayesian Kneser-Ney smoothed word n-gram language model directly from characters with no \"word\" indications. In fact, it achieves lower perplexity per character than that based on supervised segmentations. We believe this will be particularly beneficial to build a language model on such texts as speech transcripts, colloquial texts or unknown languages, where word boundaries are hard or even impossible to identify a priori.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "7" }, { "text": "For instance, TANGO algorithm (Ando and Lee, 2003) essentially finds segments such that character n-gram probabilities are maximized blockwise, averaged over n.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Note that this is different from unigrams, which are posterior distribution given data.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Imagine we try to segment an English character string \"itisrecognizedasthe\u2022 \u2022 \u2022 .\"4 Strictly speaking, this is not \"nested\" in the sense of a Nested Dirichlet process(Rodriguez et al., 2008) and could be called \"hierarchical HPYLM\", which denotes another model for domain adaptation(Wood and Teh, 2008).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "In principle fourgrams or beyond are also possible, but will be too complex while the gain will be small. For this purpose, Particle MCMC(Doucet et al., 2009) is promising but less efficient in a preliminary experiment.6 AsMurphy (2002) noted, in semi-HMM we cannot use a standard trick to avoid underflow by normalizing \u03b1[t][k] into p(k|t), since the model is asynchronous. Instead we always compute (7) using logsumexp().", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "We used different \u03bb for different word types, such as digits, alphabets, hiragana, CJK characters, and their mixtures.W is a set of words of each such type, and (13) becomes a mixture of Poisson distributions in this case.8 In the following experiments, we set a = 0.2, b = 0.1.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Notice that analyzing a test data is not easy for characterwise Gibbs sampler of previous work. Meanwhile, NPYLM easily finds the best segmentation using the Viterbi algorithm once the model is learned.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "We thank Vikash Mansinghka (MIT) for a motivating discussion leading to this research, and Satoru Takabayashi (Google) for valuable technical advice.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgments", "sec_num": null } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Exchangeability and Related Topics", "authors": [ { "first": "David", "middle": [], "last": "Aldous", "suffix": "" } ], "year": 1985, "venue": "Springer Lecture Notes in Math. 1117", "volume": "", "issue": "", "pages": "1--198", "other_ids": {}, "num": null, "urls": [], "raw_text": "David Aldous, 1985. Exchangeability and Related Topics, pages 1-198. Springer Lecture Notes in Math. 1117.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "An Estimate of an Upper Bound for the Entropy of English", "authors": [ { "first": "F", "middle": [], "last": "Peter", "suffix": "" }, { "first": "", "middle": [], "last": "Brown", "suffix": "" }, { "first": "J", "middle": [ "Della" ], "last": "Vincent", "suffix": "" }, { "first": "Robert", "middle": [ "L" ], "last": "Pietra", "suffix": "" }, { "first": "Stephen", "middle": [ "A" ], "last": "Mercer", "suffix": "" }, { "first": "Jennifer", "middle": [ "C" ], "last": "Della Pietra", "suffix": "" }, { "first": "", "middle": [], "last": "Lai", "suffix": "" } ], "year": 1992, "venue": "Computational Linguistics", "volume": "18", "issue": "", "pages": "31--40", "other_ids": {}, "num": null, "urls": [], "raw_text": "Peter F. Brown, Vincent J. Della Pietra, Robert L. Mer- cer, Stephen A. Della Pietra, and Jennifer C. Lai. 1992. An Estimate of an Upper Bound for the En- tropy of English. Computational Linguistics, 18:31- 40.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Particle Markov Chain Monte Carlo", "authors": [ { "first": "Arnaud", "middle": [], "last": "Doucet", "suffix": "" }, { "first": "Christophe", "middle": [], "last": "Andrieu", "suffix": "" }, { "first": "Roman", "middle": [], "last": "Holenstein", "suffix": "" } ], "year": 2009, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Arnaud Doucet, Christophe Andrieu, and Roman Holenstein. 2009. Particle Markov Chain Monte Carlo. in submission.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "SIGHAN Bakeoff", "authors": [ { "first": "Tom", "middle": [ "Emerson" ], "last": "", "suffix": "" } ], "year": 2005, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Tom Emerson. 2005. SIGHAN Bakeoff 2005. http://www.sighan.org/bakeoff2005/.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Markov Chain Monte Carlo in Practice", "authors": [ { "first": "W", "middle": [ "R" ], "last": "Gilks", "suffix": "" }, { "first": "S", "middle": [], "last": "Richardson", "suffix": "" }, { "first": "D", "middle": [ "J" ], "last": "Spiegelhalter", "suffix": "" } ], "year": 1996, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "W. R. Gilks, S. Richardson, and D. J. Spiegelhalter. 1996. Markov Chain Monte Carlo in Practice. Chapman & Hall / CRC.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Interpolating Between Types and Tokens by Estimating Power-Law Generators", "authors": [ { "first": "Sharon", "middle": [], "last": "Goldwater", "suffix": "" }, { "first": "Thomas", "middle": [ "L" ], "last": "Griffiths", "suffix": "" }, { "first": "Mark", "middle": [], "last": "Johnson", "suffix": "" } ], "year": 2005, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sharon Goldwater, Thomas L. Griffiths, and Mark Johnson. 2005. Interpolating Between Types and Tokens by Estimating Power-Law Generators. In NIPS 2005.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Contextual Dependencies in Unsupervised Word Segmentation", "authors": [ { "first": "Sharon", "middle": [], "last": "Goldwater", "suffix": "" }, { "first": "Thomas", "middle": [ "L" ], "last": "Griffiths", "suffix": "" }, { "first": "Mark", "middle": [], "last": "Johnson", "suffix": "" } ], "year": 2006, "venue": "Proceedings of ACL/COLING 2006", "volume": "", "issue": "", "pages": "673--680", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sharon Goldwater, Thomas L. Griffiths, and Mark Johnson. 2006. Contextual Dependencies in Un- supervised Word Segmentation. In Proceedings of ACL/COLING 2006, pages 673-680.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "A Bayesian framework for word segmentation: Exploring the effects of context", "authors": [ { "first": "Sharon", "middle": [], "last": "Goldwater", "suffix": "" }, { "first": "Thomas", "middle": [ "L" ], "last": "Griffiths", "suffix": "" }, { "first": "Mark", "middle": [], "last": "Johnson", "suffix": "" } ], "year": 2009, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sharon Goldwater, Thomas L. Griffiths, and Mark Johnson. 2009. A Bayesian framework for word segmentation: Exploring the effects of context.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Extended Viterbi algorithm for second order hidden Markov process", "authors": [ { "first": "Yang", "middle": [], "last": "He", "suffix": "" } ], "year": 1988, "venue": "Proceedings of ICPR 1988", "volume": "", "issue": "", "pages": "718--720", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yang He. 1988. Extended Viterbi algorithm for sec- ond order hidden Markov process. In Proceedings of ICPR 1988, pages 718-720.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Improving nonparameteric Bayesian inference: experiments on unsupervised word segmentation with adaptor grammars", "authors": [ { "first": "Mark", "middle": [], "last": "Johnson", "suffix": "" }, { "first": "Sharon", "middle": [], "last": "Goldwater", "suffix": "" } ], "year": 2009, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mark Johnson and Sharon Goldwater. 2009. Im- proving nonparameteric Bayesian inference: exper- iments on unsupervised word segmentation with adaptor grammars. In NAACL 2009.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Bayesian Inference for PCFGs via Markov Chain Monte Carlo", "authors": [ { "first": "Mark", "middle": [], "last": "Johnson", "suffix": "" }, { "first": "Thomas", "middle": [ "L" ], "last": "Griffiths", "suffix": "" }, { "first": "Sharon", "middle": [], "last": "Goldwater", "suffix": "" } ], "year": 2007, "venue": "Proceedings of HLT/NAACL 2007", "volume": "", "issue": "", "pages": "139--146", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mark Johnson, Thomas L. Griffiths, and Sharon Gold- water. 2007. Bayesian Inference for PCFGs via Markov Chain Monte Carlo. In Proceedings of HLT/NAACL 2007, pages 139-146.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Improved backing-off for m-gram language modeling", "authors": [ { "first": "Reinhard", "middle": [], "last": "Kneser", "suffix": "" }, { "first": "Hermann", "middle": [], "last": "Ney", "suffix": "" } ], "year": 1995, "venue": "Proceedings of ICASSP", "volume": "1", "issue": "", "pages": "181--184", "other_ids": {}, "num": null, "urls": [], "raw_text": "Reinhard Kneser and Hermann Ney. 1995. Improved backing-off for m-gram language modeling. In Pro- ceedings of ICASSP, volume 1, pages 181-184.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Building a Japanese Parsed Corpus while Improving the Parsing System", "authors": [ { "first": "Sadao", "middle": [], "last": "Kurohashi", "suffix": "" }, { "first": "Makoto", "middle": [], "last": "Nagao", "suffix": "" } ], "year": 1998, "venue": "Proceedings of LREC 1998", "volume": "", "issue": "", "pages": "719--724", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sadao Kurohashi and Makoto Nagao. 1998. Building a Japanese Parsed Corpus while Improving the Pars- ing System. In Proceedings of LREC 1998, pages 719-724. http://nlp.kuee.kyoto-u.ac.jp/nl-resource/ corpus.html.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "The Child Language Data Exchange System", "authors": [ { "first": "Brian", "middle": [], "last": "Macwhinney", "suffix": "" }, { "first": "Catherine", "middle": [], "last": "Snow", "suffix": "" } ], "year": 1985, "venue": "Journal of Child Language", "volume": "12", "issue": "", "pages": "271--296", "other_ids": {}, "num": null, "urls": [], "raw_text": "Brian MacWhinney and Catherine Snow. 1985. The Child Language Data Exchange System. Journal of Child Language, 12:271-296.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "The Infinite Markov Model", "authors": [ { "first": "Daichi", "middle": [], "last": "Mochihashi", "suffix": "" }, { "first": "Eiichiro", "middle": [], "last": "Sumita", "suffix": "" } ], "year": 2007, "venue": "NIPS", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Daichi Mochihashi and Eiichiro Sumita. 2007. The Infinite Markov Model. In NIPS 2007.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Hidden semi-Markov models", "authors": [ { "first": "Kevin", "middle": [], "last": "Murphy", "suffix": "" } ], "year": 2002, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Kevin Murphy. 2002. Hidden semi-Markov models (segment models). http://www.cs.ubc.ca/\u02dcmurphyk/ Papers/segment.pdf.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Automatic Extraction of New Words from Japanese Texts using Generalized Forward-Backward Search", "authors": [ { "first": "Masaaki", "middle": [], "last": "Nagata", "suffix": "" } ], "year": 1996, "venue": "Proceedings of EMNLP 1996", "volume": "", "issue": "", "pages": "48--59", "other_ids": {}, "num": null, "urls": [], "raw_text": "Masaaki Nagata. 1996. Automatic Extraction of New Words from Japanese Texts using General- ized Forward-Backward Search. In Proceedings of EMNLP 1996, pages 48-59.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "The Nested Dirichlet Process", "authors": [ { "first": "Abel", "middle": [], "last": "Rodriguez", "suffix": "" }, { "first": "David", "middle": [], "last": "Dunson", "suffix": "" }, { "first": "Alan", "middle": [], "last": "Gelfand", "suffix": "" } ], "year": 2008, "venue": "Journal of the American Statistical Association", "volume": "103", "issue": "", "pages": "1131--1154", "other_ids": {}, "num": null, "urls": [], "raw_text": "Abel Rodriguez, David Dunson, and Alan Gelfand. 2008. The Nested Dirichlet Process. Journal of the American Statistical Association, 103:1131-1154.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "Bayesian Methods for Hidden Markov Models", "authors": [ { "first": "L", "middle": [], "last": "Steven", "suffix": "" }, { "first": "", "middle": [], "last": "Scott", "suffix": "" } ], "year": 2002, "venue": "Journal of the American Statistical Association", "volume": "97", "issue": "", "pages": "337--351", "other_ids": {}, "num": null, "urls": [], "raw_text": "Steven L. Scott. 2002. Bayesian Methods for Hidden Markov Models. Journal of the American Statistical Association, 97:337-351.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "Semi-Supervised Structured Output Learning Based on a Hybrid Generative and Discriminative Approach", "authors": [ { "first": "Jun", "middle": [], "last": "Suzuki", "suffix": "" }, { "first": "Akinori", "middle": [], "last": "Fujino", "suffix": "" }, { "first": "Hideki", "middle": [], "last": "Isozaki", "suffix": "" } ], "year": 2007, "venue": "Proceedings of EMNLP-CoNLL", "volume": "", "issue": "", "pages": "791--800", "other_ids": {}, "num": null, "urls": [], "raw_text": "Jun Suzuki, Akinori Fujino, and Hideki Isozaki. 2007. Semi-Supervised Structured Output Learning Based on a Hybrid Generative and Discriminative Ap- proach. In Proceedings of EMNLP-CoNLL 2007, pages 791-800.", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "A Bayesian Interpretation of Interpolated Kneser-Ney", "authors": [ { "first": "Yee Whye", "middle": [], "last": "Teh", "suffix": "" } ], "year": 2006, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yee Whye Teh. 2006a. A Bayesian Interpreta- tion of Interpolated Kneser-Ney. Technical Report TRA2/06, School of Computing, NUS.", "links": null }, "BIBREF23": { "ref_id": "b23", "title": "A Hierarchical Bayesian Language Model based on Pitman-Yor Processes", "authors": [ { "first": "Yee Whye", "middle": [], "last": "Teh", "suffix": "" } ], "year": 2006, "venue": "Proceedings of ACL/COLING 2006", "volume": "", "issue": "", "pages": "985--992", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yee Whye Teh. 2006b. A Hierarchical Bayesian Lan- guage Model based on Pitman-Yor Processes. In Proceedings of ACL/COLING 2006, pages 985-992.", "links": null }, "BIBREF24": { "ref_id": "b24", "title": "A Hierarchical, Hierarchical Pitman-Yor Process Language Model", "authors": [ { "first": "Frank", "middle": [], "last": "Wood", "suffix": "" }, { "first": "Yee Whye", "middle": [], "last": "Teh", "suffix": "" } ], "year": 2008, "venue": "ICML 2008 Workshop on Nonparametric Bayes", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Frank Wood and Yee Whye Teh. 2008. A Hierarchical, Hierarchical Pitman-Yor Process Language Model. In ICML 2008 Workshop on Nonparametric Bayes.", "links": null }, "BIBREF25": { "ref_id": "b25", "title": "Bayesian Semi-Supervised Chinese Word Segmentation for Statistical Machine Translation", "authors": [ { "first": "Jia", "middle": [], "last": "Xu", "suffix": "" }, { "first": "Jianfeng", "middle": [], "last": "Gao", "suffix": "" }, { "first": "Kristina", "middle": [], "last": "Toutanova", "suffix": "" }, { "first": "Hermann", "middle": [], "last": "Ney", "suffix": "" } ], "year": 2008, "venue": "Proceedings of COLING 2008", "volume": "", "issue": "", "pages": "1017--1024", "other_ids": {}, "num": null, "urls": [], "raw_text": "Jia Xu, Jianfeng Gao, Kristina Toutanova, and Her- mann Ney. 2008. Bayesian Semi-Supervised Chi- nese Word Segmentation for Statistical Machine Translation. In Proceedings of COLING 2008, pages 1017-1024.", "links": null }, "BIBREF26": { "ref_id": "b26", "title": "An Empirical Comparison of Goodness Measures for Unsupervised Chinese Word Segmentation with a Unified Framework", "authors": [ { "first": "Hai", "middle": [], "last": "Zhao", "suffix": "" }, { "first": "Chunyu", "middle": [], "last": "Kit", "suffix": "" } ], "year": 2008, "venue": "Proceedings of IJCNLP", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hai Zhao and Chunyu Kit. 2008. An Empirical Com- parison of Goodness Measures for Unsupervised Chinese Word Segmentation with a Unified Frame- work. In Proceedings of IJCNLP 2008.", "links": null } }, "ref_entries": { "FIGREF0": { "type_str": "figure", "uris": null, "num": null, "text": "Hierarchical Pitman-Yor Language Model." }, "FIGREF1": { "type_str": "figure", "uris": null, "num": null, "text": "Chinese restaurant representation of our Nested Pitman-Yor Language Model (NPYLM)." }, "FIGREF3": { "type_str": "figure", "uris": null, "num": null, "text": "For this purpose, we maintain a forward variable \u03b1[t][k] in the bigram case. \u03b1[t][k] is the probability of a string c 1 \u2022 \u2022 \u2022 c t with the final k characters being a word (see" }, "FIGREF4": { "type_str": "figure", "uris": null, "num": null, "text": "\u03b1[0][0] = 1 and we wrote c n \u2022 \u2022 \u2022 c m as c m n . 6The rationale for (7) is as follows. Since maintaining binary variables z 1 , \u2022 \u2022 \u2022 , z N is equivalent to maintaining a distance to the nearest backward" }, "FIGREF5": { "type_str": "figure", "uris": null, "num": null, "text": "Forward filtering of \u03b1[t][k] to marginalize out possible segmentations j before t\u2212k. 1: for t = 1 to N do 2:" }, "FIGREF6": { "type_str": "figure", "uris": null, "num": null, "text": "Forward-Backward sampling of word segmentation w. (in bigram case)" }, "FIGREF8": { "type_str": "figure", "uris": null, "num": null, "text": "(b) Segmentation result. Note we used no dictionary." }, "FIGREF9": { "type_str": "figure", "uris": null, "num": null, "text": "Word segmentation of \"Alice in Wonderland \"." }, "TABREF1": { "type_str": "table", "num": null, "content": "
: Semi-supervised and supervised results. Semi-supervised results used only 10K sentences (1/5) of supervised segmentations.
", "text": "", "html": null } } } }