DTATG: An Automatic Title Generator based on Dependency Trees

Liqun Shao and Jie Wang

Department of Computer Science, University of Massachusetts, One University Avenue, 01854, Lowell, MA, U.S.A.

Keywords:

Title Generator, Central Sentence, Sentence Compression, Dependency Tree Pruning, TF-IDF.

Abstract:

We study automatic title generation for a given block of text and present a method called DTATG to generate

titles. DTATG ﬁrst extracts a small number of central sentences that convey the main meanings of the text

and are in a suitable structure for conversion into a title. DTATG then constructs a dependency tree for each

of these sentences and removes certain branches using a Dependency Tree Compression Model we devise.

We also devise a title test to determine if a sentence can be used as a title. If a trimmed sentence passes the

title test, then it becomes a title candidate. DTATG selects the title candidate with the highest ranking score

as the ﬁnal title. Our experiments showed that DTATG can generate adequate titles. We also showed that

DTATG-generated titles have higher F1 scores than those generated by the previous methods.

1 INTRODUCTION

An adequate title for a given block of text must convey

succinctly the central meanings of the text. A good ti-

tle must also be catchy. Writers would typically go

through multiple rounds of revisions to come up with

a satisfactory title. Automatic title generation (ATG)

for a given block of text aims to generate a title com-

parable to a title composed by humans. For simplicity,

we will simply call a block of text a document.

We approach ATG in two phases. In the ﬁrst phase

we search for a central sentence that is “close” to be-

ing a title in the following sense: (1) It captures the

central meanings of the text; (2) It is in a form that

can be converted into a title without too much effort.

We treat such a sentence as a title candidate. In the

second phase we compress this title candidate into the

ﬁnal title. More speciﬁcally, during the ﬁrst phase, we

extract a few central sentences that capture the main

meanings of the document. We do so using keyword-

extraction algorithms to extract keywords and rank

these sentences based on the number of keywords it

contains. During the second phase, we construct a

set of rules to select a central sentence in a suitable

form. We then construct a dependency tree for each

central sentence using a dependency parser. We trim

possible branches according to a set of empirical rules

we devise, and to form the ﬁnal title. We name our

system Dependency-Tree Automatic Title Generator

(DTATG).

If a document that already has a title, and we are

asked to generate an alternative title for the document,

then we can augment the above algorithm by generat-

ing title candidates in the ﬁrst phase as follows: After

computing central sentences, we further compute the

similarities of these sentences to the exiting title, and

select the ones with higher similarities.

ATG may also be used in applications where we

are asked to compose an article on a given title. We

may use ATG to generate a title of what we have writ-

ten and compute the similarity of the generated title

to the given title. The article would be considered on

the right track if the two titles are sufﬁciently similar.

In applications where we may need to paraphrase the

original document, we can also use ATG to generate

a new title, which may be more suitable to the new

document.

We may also use ATG to generate a label for a

topic in a document, where a topic is represented by a

set of words relevant to that topic. For example, given

a document in a corpus, we may ﬁrst use the LDA

algorithm (David M. Blei and Jordan, 2003) to iden-

tify the topics contained in the document, where each

topic is a set of words, and then use ATG to generate

a short phrase or a short sentence to label each set.

Keyword extractions and sentence compressions

are basic ATG building blocks. Keywords provide a

compact representation of the content of a document,

where a keyword is a sequence of one or more words.

Hence, a sentence having a larger number of key-

words is expected to better convey the main meanings

of a document. A popular method to identify key-

166

Shao, L. and Wang, J.

DTATG: An Automatic Title Generator based on Dependency Trees.

DOI: 10.5220/0006035101660173

In Proceedings of the 8th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2016) - Volume 1: KDIR, pages 166-173

ISBN: 978-989-758-203-5

words is to use the TF-IDF measure. Given a docu-

ment in a corpus, a term in the document with a higher

TF-IDF value implies that it appears more frequently

in the document, and less frequently in the remaining

documents. However, for a corpus of a small number

of documents, the TF-IDF value of a keyword would

almost equal to its frequency.

The Word Co-occurrence (WCO) method (Mat-

suo and Ishizuka, 2004) is a better method, which ap-

plies to a single document without a corpus. WCO

ﬁrst extracts frequent terms from the document, and

then collects a set of of word pairs co-occurring in

the same sentences (sentences include titles and sub-

titles), where one of the words is a frequent term. If

term t co-occurs frequently with a subset of frequent

terms, then it is likely to be a keyword. The authors

of WCO showed that WCO offers comparable perfor-

mance to TF-IDF without the presence of a corpus.

The Rapid Automatic Keyword Extraction

(RAKE) algorithm (Rose et al., 2010) is another

keyword extraction method based on word pair

co-occurrences. In particular, RAKE ﬁrst divides

the document into words and phrases using prede-

termined word delimiters, phrase delimiters, and

positions of stop words. RAKE then computes a

weighted graph, where each word is a node, and a

pair of words are connected with weight n if they

co-occur n times. RAKE then assigns a score to each

keyword candidate, which is the summation of scores

of words contained in the keyword. Word scores

may be calculated by word frequency, word degree,

or ratio of degree to frequency. Keyword candidates

with top scores are then selected as keywords for

the document. RAKE is superior over WCO in the

following aspects: It is simpler and achieves a higher

precision rate and about the same recall rate. We will

use RAKE to extract keywords.

Early title generation methods include Naive

Bayesian with limited vocabulary, Naive Bayesian

with full vocabulary, Term frequency and inverse doc-

ument frequency (TF-IDF), K-nearest neighbor, and

Iterative Expectation Maximization. These methods,

however, only generate an unordered set of keywords

as a title without concerning syntax. Using the F1

score metric as a base for comparisons, it was shown

through extensive experiments that the TF-IDF title

generation method has the best performance over the

other ﬁve methods (Jin and Hauptmann, 2001).

For practical purposes we would like to gener-

ate syntactically correct titles. Recent methods used

sentence trimming to convert a title candidate into a

shorter sentence or phrase, while trying to maintain

syntactic correctness. Sentence trimming has been

studied in recent years and has met with certain suc-

cess. For example, Knight and Marcu (Knight and

Marcu, 2002) and Turner and Charniak (Turner and

Charniak, 2005) used a language model (e.g., the tri-

gram model) to trim sentences. Vandegehinste and

Pan (Vandegehinste and Pan, 2004) used context-free

grammar (CFG) trees to trim a sentence to gener-

ate subtitle candidates with appropriate pronuncia-

tions for the deaf and hard-of-hearing people. They

used a spoken Dutch corpus for evaluation. More re-

cently, Zhang et al. (Zhang et al., 2013) used sentence

trimming on the Chinese text to generate titles, also

through CFG trees. We note that CFG is ambiguous

and the complexity of constructing CFG trees is high.

Dependency trees are recent development in natu-

ral language processing with a number of advantages.

For example, dependency trees offer better syntactic

representations of sentences (Sylvain, 2012) and they

are easier to work with. These advantages motivated

us to explore automatic title generation using depen-

dency trees. We present the ﬁrst such algorithm.

Our approach has the following major differences

from the previous approaches:

1. We use RAKE to generate keywords and deﬁne a

better measure to select central sentences.

2. We use dependency grammar to construct a de-

pendency tree for each title candidate for trim-

ming.

3. We construct a set of empirical rules to generate

titles.

We show that, through experiments, DTATG gener-

ates titles comparable to titles generated by human

writers. In addition to this evaluation, we also evalu-

ate the F1 scores and show that, through experiments,

DTATG is superior over the TF-IDF method (see Sec-

tion 4.3).

The remainder of this paper is organized as fol-

lows: In Section 2 we ﬁrst provide an overview of

DTATG. We then explain DTATG in detail, including

extraction of central sentences, dependency parsing,

and the dependency tree compression model. In Sec-

tion 4 we describe our experiment evaluation setups

and present results from our experiments. We con-

clude the paper in Section 5.

2 DETAIL DESCRIPTION OF

DTATG

The DTATG system framework to generate a title for

a given document is shown in Figure 1, where we as-

sume that each document already has a title for com-

parison. If a document does not have a title, then this

comparison will not be executed. Given a document,

DTATG generates a title as follows:

DTATG: An Automatic Title Generator based on Dependency Trees

167

Figure 1: DTATG system framework.

1. Extract keywords.

2. Use sentence segmentation to obtain sentences

and rank each sentence using a suitable measure

based on the number of keywords it contains. Se-

lect a ﬁxed number of sentences with the highest

rankings as the central sentences. In general, se-

lecting three central sentences would be sufﬁcient.

3. Construct a dependency tree for each central sen-

tence using a dependency parser, starting from

the sentence with the highest ranking. In par-

ticular, we use Stanford University’s open-source

tool—Stanford Dependency Parser—as our de-

pendency parser.

4. Remove certain branches of the dependency tree

based on a set of empirical rules we devise to com-

press the sentence.

5. If the trimmed sentence passes the title test we de-

vise, then output it as the title.

However, we note that not all documents have cen-

tral sentences. For example, in a document that de-

scribes a number of items to be avoided, it may simply

state that “The following items should be avoided:”

followed by a list of items. In this document, any sen-

tence that describes an item does not include the key-

word “avoided”, and the sentence that contains this

keyword does not contain any keywords for describ-

ing any item. Thus, none of the sentence in the doc-

ument can represent the central meaning of the docu-

ment, that is, none of the sentences can specify what

items should be avoided.

For convenience, we call documents with central

sentences type-1 documents and documents without

central sentences type-2 documents. While DTATG

may apply to both types of documents, it performs

better for type-1 documents.

2.1 Central Sentence Extraction

We segment each document into sentences using a

sentence-delimiter set {., ?, !, \n, :}. This set does

not include comma as compared to sentence delim-

iters used by other researchers, for we want to obtain

a complete sentence. From experience we should se-

lect only consider sentences with at most 25 words for

title candidates.

We ﬁrst use RAKE (Rose et al., 2010) to identify

keywords, where RAKE assigns each keyword with a

positive numerical score. Next we will compute the

ranking of each sentence to reﬂect the importance of

the sentence based on the keywords it contains. As-

sume that a sentence contains n keywords w

, · · · , w

and w

has a positive score s

. An obvious method to

deﬁne the rank of the sentence is to simply sum up the

scores of all the keywords contained in it as follows:

Rank

∑

i=1

. (1)

However, we observe that using Rank

we may end up

giving a sentence containing a larger number of key-

words of lower scores a higher rank than a sentence

containing fewer keywords of higher scores. For ex-

ample, suppose that sentence S

contains two key-

words with scores of (4, 5) and sentence S

contains

four keywords with scores of (2, 2, 3, 3). Then S

has

a higher Rank

ranking than S

. This contradicts to

the common sense that S

would be more relevant to

the central meanings.

Thus, we would want keywords of higher scores to

carry more weights, so that the ranking of a sentence

with smaller number of keywords of much higher

scores is higher than the ranking of a sentence with

KDIR 2016 - 8th International Conference on Knowledge Discovery and Information Retrieval

168

larger number of keywords of much lower scores. To

achieve this, we use the power of 2 to amplify the

scores, giving rise to the following measure:

Rank

∑

i=1

. (2)

In the example above, it is easy to see that S

has a higher Rank

ranking. On the other hand, if

contains three keywords each with a score of 4,

then we would consider S

and S

equally important.

The Rank

measure ensures this effect. Moreover, if

contains four keywords with scores of (1, 4, 4, 4),

then we would want S

to have a higher ranking than

. Again, the Rank

measure also ensures this effect.

Empirical results indicate that Rank

is a better

measure for ranking sentences. For example, Table

1 shows two central sentences obtained using Rank

and Rank

from a news article as an example, both

having the highest score under the respective mea-

sure. It is evident that, compared with the origi-

nal title, the central sentence obtained using Rank

presents a better choice.

Table 1: Examples of central sentence extraction.

Original title Bad e-mail habits sustains spam

Rank

People must resist their basic

instincts to buy from spam mails

Rank

One in ten users have bought

products advertised in junk mail

DTATG uses Rank

to select central sentences as

follows: Order sentences ﬁrst by rank and then by

the sentence length. In other words, we ﬁrst select

the sentence with the highest rank and if two sen-

tences have the highest rank, we select the shorter

one. Our experiments indicate that this method of se-

lecting central sentences improves performance.

The sentence we selected may have multiple

clauses separated by commas. If the sentence has

more than two clauses, we keep the two clauses that

have the highest ranking and separate them using a

space. This makes it easier to perform sentence com-

pression in the next step.

To generate suitable sentences for trimming, we

apply the following empirical rules to avoid inappro-

priate sentences.

• Discard sentences that contain a pronoun subject.

• Discard sentences that contain predicate verbs in

{am, is, are}.

2.2 Dependency Trees

We use the Stanford Dependency Parser (SDP),

available at http://nlp.stanford.edu/software/stanford-

dependencies.shtml, to obtain grammatical relations

between words in a sentence. The relations are pre-

sented in triplets:

(relation, governor, dependent).

The tree is constructed by the relations be-

tween each word in the sentence. For exam-

ple, nsubj stands for nominal subject, which

is the syntactic subject of a clause. Univer-

sal Dependencies are documented online at

http://universaldependencies.org/u/dep/index.html.

For example, Figure 2 depicts a dependency

tree for the sentence “Microsoft warned PC

users to update their systems”. In addition, we

may use the Stanford Parser (SP), available at

http://nlp.stanford.edu/software/lex-parser.shtml, to

obtain part-of-speech tagging.

Figure 2: The dependency tree of sentence “Microsoft

warned PC users to update their systems”.

2.3 Dependency Tree Compression

Model

We devise a dependency tree compress model

(DTCM) to generate titles. First, we deﬁne the fol-

lowing set of empirical rules to specify what cannot

be trimmed. These rules were formed based on our

experiences from working with a large number of text

documents. We refer to these rules as the to-be-kept

rules.

1. If the relation is nsubj or nsubjpass, then keep

both the governor and dependent.

2. If the relation is dobj or iobj, then keep both the

governor and dependent.

3. If the relation is compound, then keep both the

governor and dependent.

4. If the relation is root, then keep the dependent.

5. If the relation is nmod, then keep the dependent

and the preposition in nmod.

DTATG: An Automatic Title Generator based on Dependency Trees

169

6. If the relation is nummod, then keep both the gov-

ernor and dependent.

We note that the remainder of the sentence after

removing all the other relations not speciﬁed the to-

be-kept rules can still be understandable by humans.

DTCM proceeds as follows:

1. Traverse the entire dependency tree in preorder.

2. For each leaf node t, if it is not a keyword or a part

of the keyword or the relation of its parent edge

is not in the set of to-be-kept rules, then remove

node t.

In general, DTCM deletes all the unnecessary

branches; most of these branches would have rela-

tions such as det, aux, and amod.

Figure 3: The trimmed dependency tree using DTCM for

sentence “Market concerns about the deﬁcit has hit the

greenback.”

Figure 3 shows a trimmed dependency tree us-

ing DTCM for a central sentence “Market concerns

about the deﬁcit has hit the greenback.” The output is

“Market concerns about deﬁcit hit greenback”, which

would be a suitable title.

2.4 Compression Rate

To transform a trimmed dependency tree into a title,

we output the words in the same order of the original

sentence. Let y be a compressed sentence for a sen-

tence x = w

· · · w

, where w

is a word. We use

“0” to mark a word to be removed and “1” to mark a

word to be retained. Then a compression rate r for a

sentence is deﬁned as follows:

r =

# of 1s in y

. (3)

If r is too large or too small, we may delete more or

fewer words, respectively, to make the output more

reasonable. For example, since the length of a good

title should not exceed 10 words (this is a common-

sense rule), if a central sentence is more than 20 words

and the compression rate r is less than 50% after prun-

ing, then we may apply DTCM again on the trimmed

sentence with the hope to obtain a higher compression

rate.

The following is a set of empirical rules for dele-

tion and we refer to them as to-be-deleted rules. We

may keep adding rules to this set.

• Delete the ﬁrst adverbial phrase and the last ad-

verbial phrase.

• Delete “X says” and “X said”, where X is a noun

or pronoun.

• If there are two clauses connected with “and” and

the ﬁrst clause has subject and verb, delete the sec-

ond clause with “and”.

• If there is a clause starting with “that” and the

clause has a subject and a verb, then delete all the

words before “that” (including “that”).

• If there are more than one nmod relations next to

each other, then delete all the nmod relations ex-

cept the last one.

3 TITLE TEST

A good title must pass the title test, which consists of

the following three individual tests.

The Conciseness Test.

• A title should not exceed 15 words.

• A title must not have clauses.

• A title should have the following structure: “Sub-

ject + Verb + Object” or “Subject + Verb”. Sub-

ject must be speciﬁc: it can be a noun but not a

pronoun.

The Fluency Test. A title should contain no gram-

matical errors.

The Topic-relevance Test. A title must convey at

least one major meaning of the document.

DTATG uses central sentences to achieve topic

relevance, and uses dependency grammar and empir-

ical rules to achieve conciseness and ﬂuency. We will

use the title test to evaluate the quality of the gener-

ated titles by DTATG.

4 EVALUATIONS

We evaluate DTATG on a corpus of English

documents obtained from the BBC news web-

site (Greene and Cunningham, 2006). The dataset

can be found at http://mlg.ucd.ie/datasets/bbc.html

and https://drive.google.com/open?id=0B_-

6yYneQ1_8Ml91TnY0THJoYXc, which consists

of 2,225 documents in ﬁve topical areas from the

year of 2004 to the year of 2005. These ﬁve topical

KDIR 2016 - 8th International Conference on Knowledge Discovery and Information Retrieval

170

areas are business, entertainment, politics, sport, and

tech. The corpus provides raw text ﬁles with original

articles and titles. The datasets include type-1 and

type-2 documents. In our experiments, we will use

type-1 articles in the business, entertainment, politics,

and tech categories and type-2 articles in the sport

category.

4.1 Experiment Setup

We carried out experiments on a computer with 2.7

GHz dual-core Intel Core i5 CPU and 8 GB memory.

In our experiments, we used a Python implementation

of the RAKE algorithm to extract keywords, and

the Stanford Parser to generate dependency trees,

where the source code for implementing RAKE can

be found at https://github.com/aneesha/RAKE

and the Stanford Parser can be downloaded

from http://nlp.stanford.edu/software/lex-

parser.shtml#Download.

4.2 Results and Discussions

We randomly selected 50 articles in each category to

evaluate. We asked English speakers to rank each

DTATG-generated title in the category of conciseness,

ﬂuency, and topic relevance using a 5-point scale.

Rating Deﬁnition on Conciseness (CON)

5: Excellent; every word in the sentence is to the

point that makes a perfect title.

4: Very good; the sentence contains no redundant

words that makes a very good title.

3: Good; the sentence has some redundant words,

but may still be used as a title.

2: Weak; the sentence has some redundant words,

but is a poor choice as a title.

1: Poor; the sentence contains a lot of redundant

words and should not be used as a title at all.

Rating Deﬁnition on Fluency (FLU)

5: Excellent; the sentences contains no grammatical

ﬂaws.

4: Very good; the sentences contains only minor

grammatical ﬂaws.

3: Good; the sentence contains some grammatical

ﬂaws.

2: Weak; the sentence is only partially intelligible,

with serious grammatical ﬂaws.

1: Poor; the sentence makes no sense at all.

Rating Deﬁnition on Topic Relevance (TR)

5: Excellent; the sentence has perfect topic rele-

vancy.

4: Very good; the sentence has signiﬁcant topic rele-

vancy.

3: Good; the sentence has moderate topic relevancy.

2: Weak; the sentence has marginal topic relevancy.

1: Poor; the sentence is topic irrelevant.

Evaluation Results

Each evaluation category was evaluated by 5 human

evaluators. We then average their scores. On the con-

ciseness and ﬂuency evaluation, we asked the eval-

uators to compare the original titles with DTATG-

generated titles. On the topic relevance evaluation,

we asked the evaluators to make a subjective judg-

ment on how much the DTATG-generated titles repre-

sent the main topics of the article. For this evaluation,

the evaluators needed to read the original articles. We

note that most articles in the sport category are type-2

documents. In these documents we simply used the

ﬁrst sentence as the central sentence.

To better assess the quality of the DTATG results,

we also ask evaluators to evaluate the original titles

using the same rules, and we use the original as the

comparison baseline. Results are given in Table 2

and Figure 4 for a better visual. In Figure 4, the dot-

ted lines represent the scores of titles generated by

DTATG and the solid lines represent the scores of

original titles.

Figure 4: Line graph of DTATG-generated titles and origi-

nal titles on CON, FLU, and TR.

From Table 2 we can see that the titles generated

by DTATG have scores over 4 on ﬂuency and topic

relevance on all categories. Moreover, most of the

titles are concise except in the tech category. The

reason why the generated titles for the tech articles

DTATG: An Automatic Title Generator based on Dependency Trees

171

Table 2: The average scores for DTATG-generated titles compared with the original titles.

Category Base-CON CON Base-FLU FLU Base-TR TR

Business 4.84 4.40 4.76 4.68 4.72 4.12

Entertainment 4.80 4.20 4.68 4.72 4.76 4.16

Politics 4.72 4.16 4.76 4.52 4.68 4.04

Sport 4.80 4.12 4.60 4.44 4.40 4.68

Tech 4.96 3.84 4.44 4.32 4.32 4.12

Table 3: Comparison examples of DTATG-generated titles with the original title showing the score (CON, FLU, TR).

Original title Blogs take on the mainstream (5, 5, 4)

Central sentence The blogging movement has been building up for many years.

DTATG title Blogging movement building up for years (5, 5, 5)

Remark The DTATG title contains more information, and so we consider it better than

the original title.

Original title Day-Lewis set for Berlin honour (5, 5, 4)

Central sentence Japan’s oldest ﬁlm studio will also be honoured along with Day-Lewis.

DTATG title Japan’s oldest ﬁlm studio honoured with Day-Lewis (5, 5, 5)

Remark The original article talked about both Day-Lewis’s and Japan’s oldest ﬁlm studio’s

contributions. The original title only mentioned Day-Lewis. The DTATG title

mentioned both, and so we would consider it a better title.

Original title Scots smoking ban details set out (5, 5, 5)

Central sentence A comprehensive ban on smoking in all enclosed public places in Scotland.

DTATG title Comprehensive ban on smoking in public places in Scotland (5, 5, 5)

Remark The DTATG title is as good as the original title made by a professional writer.

Original title Xbox 2 may be unveiled in summer (5, 5, 5)

Central sentence Since its launch, Microsoft has sold 19.9 million units worldwide.

DTATG title Microsoft sold 19.9 million units worldwide (5, 5, 4)

Remark The DTATG title is missing Xbox as the name of the units sold, and so it is not

a very good title.

are not as concise is because these articles often con-

tain long technology names or long company names,

which make titles longer. Nevertheless, DTATG in

general does reasonably well on all evaluation cate-

gories of conciseness, ﬂuency, and topic relevance.

DTATG-generated titles are understandable and ac-

ceptable by human evaluators. From Figure 4, we can

easily tell that titles generated by DTATG are close

to baselines. Titles of FLU in Entertainment and TR

in Sport categories are better than the baseline, which

means they are even better than the original titles be-

cause they provide more information (see Table 3).

Detailed information of these examples, in-

cluding the original title, the DTATG-generated

title, the original text, the keywords, the central

sentence, the dependency tree for the central sen-

tence, and the pruned branches, can be found at

http://119.29.118.23/title_generator/index.html.

We also implemented a system at

http://www.cwant.cn/title_en/ that allows users

to enter text, upload a txt ﬁle, or enter an URL to

generate a title automatically using DTATG.

We note that the performance of DTATG is con-

ﬁned by the choice of central sentences. To produce

a good title, we must have a good central sentence to

begin with. For example, in the last example in Table

3, if “Xbox” is included in the central sentence to read

”Since its launch, Microsoft has sold 19.9 million

Xbox units worldwide”, then the DTATG-generated

title would have been “Microsoft sold 19.9 million

Xbox units worldwide”, which would be a good ti-

tle. However, the original article mainly talked about

the success of Xbox and only mentioned brieﬂy that

Xbox 2 would be available in May. Thus, no central

sentences would be able to capture Xbox 2.

4.3 Comparision with TF-IDF

For title generation using the TF-IDF method, the

words in the document with highest TF-IDF would be

chosen for the title. Given a document, let T1 denote

the title generated by DTATG and T2 the original title.

Let precision denote the number of identical words in

T1 and T2, divided by the number of words in T1; and

KDIR 2016 - 8th International Conference on Knowledge Discovery and Information Retrieval

172

recall the number of identical words in T1 and T2, di-

vided the number of words in T2. Then the F1 score

is deﬁned by

F1 =

2 × precision × recall

precision + recall

. (4)

We randomly selected 100 articles from datasets in

Section 4 and computed the F1 score for each article.

The average F1 scores are shown in Figure 5.

Figure 5: Comparison of DTATG and TF-IDF on the F1

score.

In four out of ﬁve categories, the titles generated

by DTATG perform better than those generated by

TF-IDF. In particular, for the category of sport, the

average F1 score under DTATG is much higher than

that under TF-IDF. In the category of politics, the av-

erage F1 score under DTATG is slightly lower than

that under TF-IDF.

5 CONCLUSION

We presented DTATG for automatic title generation

on a given block of text. DTATG is a system com-

bining central sentence selection and dependency tree

pruning. DTATG is unsupervised and can generate

titles quickly and syntactically. DTATG, however, is

conﬁned by the choice of central sentences. We plan

to study how to combine several central sentences and

compress them using deletion, substitution, reorder-

ing, and insertion to generated better titles.

ACKNOWLEDGEMENTS

This work was supported in part by a research grant

from Wantology LLC. We thank Yiqi Bai, Shan

(Ivory) Lu, Jing Ni, Jingwen (Jessica) Wang, Hao

Zhang, and a few other people for helping us evalu-

ate the titles generated by DTATG.

REFERENCES

David M. Blei, A. Y. N. and Jordan, M. I. (2003). Latent

dirichlet allocation. The Journal of machine Learning

research, 3(2003): 993–1022.

D. Greene and P. Cunningham. (2006). Practical solutions

to the problem of diagonal dominance in kernel docu-

ment clustering. In ICML.

Kevin Knight and Daniel Marcu. (2002). Summarization

beyond sentence extraction: A probabilistic approach

to sentence compression. In Artiﬁcial Intelligence,

139(1): 91-107.

Rong Jin and Alexander G. Hauptmann. (2001). Automatic

Title Generation for Spoken Broadcast News. In Pro-

ceedings of HLT-01, 2001, pp. 1

lC3.

Stuart Rose, Dave Engel, Nick Cramer and Wendy Cow-

ley. (2010). Automatic keyword extraction from in-

dividual documents. In Text mining applications and

theory, 2010, pp. 3-19.

Jenine Turner and Eugene Charniak. (2005). Supervised

and unsupervised learning for sentence compression.

In Proceedings of the 43rd Annual Meeting of the As-

sociation for Computational Linguistics, Ann Arbor,

Mich., 25-30 June 2005, pp. 290-297.

Vincent Vandegehinste and Yi Pan. (2004). Sentence com-

pression for automated subtitling: A hybrid approach.

In Proceedings of the ACL-04, 2004 (pp. 89-95).

Y. Matsuo and M. Ishizuka. (2004). Keyword extraction

from a single document using word co-occurrence sta-

tistical information. International Journal on Artiﬁcial

Intelligence Tools, 13(1): 157-169.

Sylvain Kahane. (2012). Why to choose dependency rather

than constituency for syntax: a formal point of view.

In Meanings, Texts, and other exciting things: A

Festschrift to Commemorate the 80th Anniversary of

Professor Igor A. Mel’

cuk, Languages of Slavic Cul-

ture, Moscou, pp. 257-272.

Yonglei Zhang, Cheng Peng, and Hongling Wang. (2013).

Research on Chinese Sentence Compression for

the Title Generation. Chinese Lexical Semantics,

edited by Xiao Guozheng & Ji Donhong. Heideberg:

Springer.

DTATG: An Automatic Title Generator based on Dependency Trees

173