This paper deals with an application allowing the automatic titling of texts. This one consists of four stages: Corpus acquisition, candidate sentence determination for the titling, extraction of noun phrases among the candidate sentences, and nally the choice of the title.
In this paper, we present an application dealing with an automatic approach providing a title to a document, which meets the di erent characteristics of human issued titles. So, when a title is absent, for instance in emails without objects or to determine the le title for saving, the described method enables the user to save time by informing him/her about the content in a single glance. In addition, it is designed to meet at least one of the criteria of the standard W3C. Indeed, titling web pages is one of key elds of the web page accessibility, such as de ned by associations for the disabled. The goal is to enhance the page readability. Moreover, a relevant title is an important issue for the webmaster improving the indexation of web pages. Let us note that titling is not a task to be confused with automatic summarization, text compression, and indexation, although it has several common points with them. This will be detailed in the 'related work' section.
While a lot of applications are borned in the NLP domain,
it seems that no application was realized to title
automatically textual documents. As for articles, it was noticed that
elements appearing in the title are often present in the body
of the text [
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee.
Copyright 20XX ACM X-XXXXX-XX-X/XX/XX ...$10.00. 3. 3.1
A title is not exactly the smallest possible abstract. While
a summary, the most condensed form of a text, has to give an
outline of the text contents that respects the text structure,
a title indicates the treated subject in the text without
revealing all the content. Text compression could be
interesting for titling if a strong compression could be undertaken,
resulting in a single relevant word group. Compression texts
methods (e.g. [
A title is not an index : A title does not necessarily contain key words (and indexes are key words), and might present a partial or total reformulation of the text (what an index is not).
Finally, a title is thus a full entity, has own functions, and titling has to be sharply distinguished from summarizing and indexing.
The global process in order to automatically title a document is composed of the following steps:
Step 0: Corpus Acquisition: Determining the
characteristics of the texts to be titled;
Step 1: Candidate Sentences Determination. We
assume that any text contains at least a few sentences
that would provide the relevant sentence for titling. In
this article, we suppose that the terms used in the title
can be located in the rst sentences of the text [
Step 3: Selecting a Title by the ChTitres Approach
based on the use of TF-IDF measure [
The application is available in English and French: http://www.lirmm.fr/ lopez/. This online application, developped with HTML, PHP, and MySQL has an user-friendly interface. On the screen shot (see Fig. 1), all parts of the interface are annotated with a letter.
A: It enables to choose which method the user can apply in order to title a given document (TMAX, TSUM, and other methods). By clicking the name of the method, the user nds the explanations about it.
B: The interface enables to choose which method will be apply for subtitling.
C: The text area has to receive the text to title. The text block must be separated by 'carriage returns' in order to be subtitled.
D: Some application examples (specialized texts) are proposed in a list.
E: Link allowing to pass in French mode.
The result page returns all the titles (and subtitles) according to the chosen methods on the interface page.
Ten experts have evaluated our methods (240 titles have been evaluated with 3 corpora). For every text, six titles were suggested1 among all the titles determined according to the methods TMAX and TSUM , as well as the real title T R. Three other titles (A1; A2; A3) are exposed in a random way from the list of noun phrases extracted among those that were rejected by the process. Comparing the evaluation of 1Identical titles obtained with di erent approaches are not given. rejected NP with selected ones will allow, in particular, the estimation of the selection process accuracy.
For every 'candidate' title, the expert has to apreciate its relevance to the document contents with the following scale: Very relevant (C1), Relevant (C2), I don't know (C3), not very relevant (C4), not relevant at all (C5). For each of these Cn judgements, a digital value is assigned: 2 for C5, 1 for C4, 0 for C3, +1 for C2 and+2 for C1. The nal note obtained for a title is the mean value of the experts given grades (see Fig. 2). So, the higher the value, the more accurate the NP, as a title.
Based on this expert evaluation, the experiments show that e-mails titles returned by our tool are relevant (0<score<1, see Table 1). In particular, evaluation results show that automatic e-mail titles (0.61) are more relevant than real titles (0.57) (with TMAX ).
CORPORA TR E-mails 0.57 Fora 1.15 Mailing lists 1.8
TMAX 0.61 0.42 0.81
TSUM 0.52 0.58 0.43 4.
The experts have con rmed that the titles built by our automatic titling tool are relevant. It is a possible bene t of an automatic method that might build a more relevant title than a 'real' one, and is a time saving procedure for a heavy e-mails writer.