A Generic Sentence Trimmer with CRFs
Nomoto, Tadashi

Article Structure

Abstract

The paper presents a novel sentence trimmer in Japanese, which combines a non-statistical yet generic tree generation model and Conditional Random Fields (CRFs), to address improving the grammaticality of compression while retaining its relevance.

Introduction

For better or worse, much of prior work on sentence compression (Riezler et al., 2003; McDonald, 2006; Turner and Charniak, 2005) turned to a single corpus developed by Knight and Marcu (2002) (K&M, henceforth) for evaluating their approaches.

A Sentence Trimmer with CRFs

Our idea on how to make CRFs comply with grammar is quite simple: we focus on only those label sequences that are associated with grammatically correct compressions, by making CRFs look at only those that comply with some grammatical constraints G, and ignore others, regardless of how probable they are.1 But how do we find compressions that are grammatical?

Features in CRFs

We use an array of features in CRFs which are either derived or borrowed from the taxonomy that a Japanese tokenizer called JUMAN and KNP,6 a Japanese dependency parser (aka Kurohashi-Nagao Parser), make use of in characterizing the output they produce: both JUMAN and KNP are part of the compression model we build.

The Dependency Path Model

In what follows, we will describe somewhat in detail a prior approach to sentence compression in Japanese which we call the ”dependency path model,” or DPM.

Evaluation Setup

We created a corpus of sentence summaries based on email news bulletins we had received over five to six months from an online news provider called Nikkei Net, which mostly deals with finance and politics.9 Each bulletin consists of six to seven news briefs, each with a few sentences.

Results and Discussion

We ran DPM and GST on NICOM in the 10-fold cross validation format where we break the data into 10 blocks, use 9 of them for training and test on the remaining block.

Conclusions

This paper introduced a novel approach to sentence compression in Japanese, which combines a syntactically motivated generation model and CRFs, in or-

Topics

CRFs

Appears in 15 sentences as: CRFs (17) CRFs’ (1)
In A Generic Sentence Trimmer with CRFs
  1. The paper presents a novel sentence trimmer in Japanese, which combines a non-statistical yet generic tree generation model and Conditional Random Fields ( CRFs ), to address improving the grammaticality of compression while retaining its relevance.
    Page 1, “Abstract”
  2. What sets this work apart from them, however, is a novel use we make of Conditional Random Fields ( CRFs ) to select among possible compressions (Lafferty et al., 2001; Sutton and McCallum, 2006).
    Page 1, “Introduction”
  3. An obvious benefit of using CRFs for sentence compression is that the model provides a general (and principled) probabilistic framework which permits information from various sources to be integrated towards compressing sentence, a property K&M do not share.
    Page 1, “Introduction”
  4. Nonetheless, there is some cost that comes with the straightforward use of CRFs as a discriminative classifier in sentence compression; its outputs are often ungrammatical and it allows no control over the length of compression they generates (Nomoto, 2007).
    Page 1, “Introduction”
  5. We tackle the issues by harnessing CRFs with what we might call dependency truncation, whose goal is to restrict CRFs to working with candidates that conform to the grammar.
    Page 1, “Introduction”
  6. Our idea on how to make CRFs comply with grammar is quite simple: we focus on only those label sequences that are associated with grammatically correct compressions, by making CRFs look at only those that comply with some grammatical constraints G, and ignore others, regardless of how probable they are.1 But how do we find compressions that are grammatical?
    Page 2, “A Sentence Trimmer with CRFs”
  7. 1Assume as usual that CRFs take the form, p(le) 0< eX10 21m- )‘jfj(yk7yk—17X)+ Z,- Migi($k7ykax))
    Page 2, “A Sentence Trimmer with CRFs”
  8. (1) fj and g, are ‘features’ associated with edges and vertices, respectively, and k: E C, where C denotes a set of cliques in CRFs .
    Page 2, “A Sentence Trimmer with CRFs”
  9. Finally, in order for CRFs to work with the compressions, we need to translate them into a sequence of binary labels, which involves labeling an element token, bunsetsu or a word, with some label, e.g., 0 for ’remove’ and l for ‘retain,’ as in Figure 6.
    Page 4, “A Sentence Trimmer with CRFs”
  10. Because y4 is not part of G(S), it is not considered a candidate for a compression for y, even if its likelihood may exceed those of others in G We note that the approach here does not rely on so much of CRFs as a discriminative classifier as CRFs as a strategy for ranking among a limited set of label sequences which correspond to syntactically plausible simplifications of input sentence.
    Page 4, “A Sentence Trimmer with CRFs”
  11. We use an array of features in CRFs which are either derived or borrowed from the taxonomy that a Japanese tokenizer called JUMAN and KNP,6 a Japanese dependency parser (aka Kurohashi-Nagao Parser), make use of in characterizing the output they produce: both JUMAN and KNP are part of the compression model we build.
    Page 5, “Features in CRFs”

See all papers in Proc. ACL 2008 that mention CRFs.

See all papers in Proc. ACL that mention CRFs.

Back to top.

sentence compression

Appears in 8 sentences as: sentence compression (8)
In A Generic Sentence Trimmer with CRFs
  1. For better or worse, much of prior work on sentence compression (Riezler et al., 2003; McDonald, 2006; Turner and Charniak, 2005) turned to a single corpus developed by Knight and Marcu (2002) (K&M, henceforth) for evaluating their approaches.
    Page 1, “Introduction”
  2. Despite its limited scale, prior work in sentence compression relied heavily on this particular corpus for establishing results (Turner and Charniak, 2005; McDonald, 2006; Clarke and Lapata, 2006; Galley and McKeown, 2007).
    Page 1, “Introduction”
  3. An obvious benefit of using CRFs for sentence compression is that the model provides a general (and principled) probabilistic framework which permits information from various sources to be integrated towards compressing sentence, a property K&M do not share.
    Page 1, “Introduction”
  4. Nonetheless, there is some cost that comes with the straightforward use of CRFs as a discriminative classifier in sentence compression ; its outputs are often ungrammatical and it allows no control over the length of compression they generates (Nomoto, 2007).
    Page 1, “Introduction”
  5. In the context of sentence compression , a linear programming based approach such as Clarke and Lapata (2006) is certainly one that deserves consideration.
    Page 2, “A Sentence Trimmer with CRFs”
  6. 2Note that a sentence compression can be represented as an array of binary labels, one of them marking words to be retained in compression and the other those to be dropped.
    Page 2, “A Sentence Trimmer with CRFs”
  7. In what follows, we will describe somewhat in detail a prior approach to sentence compression in Japanese which we call the ”dependency path model,” or DPM.
    Page 5, “The Dependency Path Model”
  8. This paper introduced a novel approach to sentence compression in Japanese, which combines a syntactically motivated generation model and CRFs, in or-
    Page 8, “Conclusions”

See all papers in Proc. ACL 2008 that mention sentence compression.

See all papers in Proc. ACL that mention sentence compression.

Back to top.

Dependency path

Appears in 4 sentences as: Dependency path (1) dependency paths (1) ‘Dependency Path (1) ”dependency path (1)
In A Generic Sentence Trimmer with CRFs
  1. Later in the paper, we will introduce an approach called the ‘Dependency Path Model’ (DPM) from the previous literature (Section 4), which purports to provide a robust framework for sentence compres-
    Page 1, “Introduction”
  2. We begin by locating terminal nodes, i.e., those which have no incoming edges, depicted as filled circles in Figure 3, and find a dependency (singly linked) path from each terminal node to the root, or a node labeled ‘E’ here, which would give us two paths p1 = ACDE and p2 = BCDE (call them terminating dependency paths , or TDPs).
    Page 3, “A Sentence Trimmer with CRFs”
  3. In what follows, we will describe somewhat in detail a prior approach to sentence compression in Japanese which we call the ”dependency path model,” or DPM.
    Page 5, “The Dependency Path Model”
  4. Dependency path length (DL) refers to the number of (singly linked) dependency relations (or edges) that span two bunsetsa’s.
    Page 6, “The Dependency Path Model”

See all papers in Proc. ACL 2008 that mention Dependency path.

See all papers in Proc. ACL that mention Dependency path.

Back to top.

gold standard

Appears in 4 sentences as: gold standard (4)
In A Generic Sentence Trimmer with CRFs
  1. The relevance test, on the other hand, consisted of paired compressions along with the associated gold standard compressions.
    Page 7, “Results and Discussion”
  2. We randomly picked 200 of them from NICOM-g, at each compression rate, and asked the participants to make a subjective judgment on how much of the content in a compression semantically overlap with that of the gold standard , on a scale of 1 to 5 (Table 3).
    Page 7, “Results and Discussion”
  3. Also included in the survey are 200 gold standard compressions, to get some idea of how fluent “ideal” compressions are, compared to those generated by machine.
    Page 7, “Results and Discussion”
  4. Since the average CR of gold standard compressions was 60%, we report their fluency at that rate only.
    Page 8, “Results and Discussion”

See all papers in Proc. ACL 2008 that mention gold standard.

See all papers in Proc. ACL that mention gold standard.

Back to top.

generation model

Appears in 3 sentences as: generation model (2) generation models (1)
In A Generic Sentence Trimmer with CRFs
  1. The paper presents a novel sentence trimmer in Japanese, which combines a non-statistical yet generic tree generation model and Conditional Random Fields (CRFs), to address improving the grammaticality of compression while retaining its relevance.
    Page 1, “Abstract”
  2. To address the issue, rather than resort to statistical generation models as in the previous literature (Cohn and Lapata, 2007; Galley and McKeown, 2007), we pursue a particular rule-based approach we call a ‘dependency truncation,’ which as we will see, gives us a greater control over the form that compression takes.
    Page 2, “A Sentence Trimmer with CRFs”
  3. This paper introduced a novel approach to sentence compression in Japanese, which combines a syntactically motivated generation model and CRFs, in or-
    Page 8, “Conclusions”

See all papers in Proc. ACL 2008 that mention generation model.

See all papers in Proc. ACL that mention generation model.

Back to top.