Computer-assisted data-driven text analysis with TATOE

1. Introduction
Data-driven and corpus-based text analysis involves considerably variant tasks and aims. For natural language processing (NLP) applications text
analysis practice includes tasks ranging from
knowledge acquisition for dictionary construction, grammar development, testing and validation, to discourse analysis for automatic text understanding or/and text generation.
Text analysts in computational linguistics form,
thus, a group with mixed objectives. However,
although their ultimate aims of analysis may differ, initial exploration, modeling and codification
of the data and subsequent extraction of knowledge which may be dependent on the already performed categorization and codification are common
tasks. Analysis typically covers more than one
level of linguistic description whereby information obtained from one level is used in order to
gain a deeper understanding of another and contributing, thus, to a deeper understanding of a
single piece of discourse.
In corpus-based NLP, text analysts work with
corpora whose size may vary from some thousands of words to tens of millions. The texts can
be monolingual or bi- or multi-lingual. When
more languages than one are considered the corpora can be parallel or non-parallel. Analysis may
concentrate on single words, word collocations,
word groups (syntactic or semantic units) clauses,
sentences, paragraphs, and so on. Furthermore –
and as one of the main reasons for adopting a
corpus-based approach to analysis – the focus of
analysis is the language use in (real) context.
Computational support is required for the processing of machine-readable monolingual or multilingual corpora, the integration of programs for
morphological analysis, Part of Speech (PoS) annotation, lemmatization, etc., extracting statistical
information concerning frequency of occurrence
of word tokens, word types or lemmas and on-text
selection of a word or string of words and presentation of all of its occurrences in context. When it
comes to working intellectually with the texts,
support is further needed for assigning text segments to categories belonging to schemata which
are hierarchically or non-hierarchically structured.
This necessitates support for (i) importing a schema which is to be used for codification and classification and (ii) support for defining and constructing one’s own schema. Also, functionalities
are required which enable merging, modifying and
deleting categories of the schema (or schemata)
used. Texts may already come annotated (SGML,
morphological, syntactic, etc.) and means for integrating and re-using this annotation are essential.
Search for certain linguistic cues, frequency of
occurrence information, annotation, representation and retrieval of the annotated information are
all tasks which need to be performed in text analysis. Currently, fully automated annotation procedures are possible for some linguistic levels of
description, for instance morphological or PoS
tagging, but not for all. A way to fill the gap
between the knowledge which is acquired in a
fully automated manner and knowledge which is
still required for text analysis is possible by means
of semi-automated “mark up”.
2. Text analysis – the work context
Over the last few years we have performed intensive corpus-based text analysis of corpora belonging to different domains, in different languages
and, also, with different analysis objectives:
• sublanguage analysis: (Alexa, 1993): analysis of English and Greek job advertisement
texts for the purposes of multilingual text
generation systems
• genre and register analysis (Alexa, 1995):
analysis of different text types – artists’ biographies and archeological site descriptions
– typical of encyclopedic reference works in
order to specify contextual (text structure)
choices which are to guide text type sensitive
automatic generation in the Komet-PenmanMultiLingual (KPML) text generation system (Bateman 1996), (Teich and Bateman
1994)
• semi-automatic information extraction: analysis of a corpus of English biography texts
in order to first identify prominent and text
type specific event structures, i.e. main verbs
together with the concepts (their participants
and the temporal and spatial information),
11
parse these events and update the knowledge
base used for the Editor’s Workbench application (for the general context of this work
see Rostek et al., 1994 and Rostek and Moehr, 1994)
• domain specific thesaurus construction: analysis of press release messages issued by the
Deutsche Presse Agentur (DPA) for the construction of a semantic network to support
fact extraction from these messages
• thematic development in encyclopedic reference texts: exploration, determination and
subsequent specification of thematic development preferences typical of English artist
biographies and archeological site descriptive texts in order to specify the text type
specific thematic choices for automatic text
generation.
For all the above objectives computational support
has been required to structure, compile and edit the
text data and to further annotate parts of the text
corpora and extract linguistically variant
knowledge.
Aiming at supporting such text analysis tasks as
the ones listed above we have designed and implemented TATOE, a Text Analysis Tool with Object
Encoding, the functionalities of which we present
in this paper. Section 3 lists specific requirements
for supporting computationally corpus-based text
analysis and presents some of the current tools for
this task. Section 4 describes TATOE and its main
functionalities. Section 5 provides the technical
characteristics of the tool. Finally, section 6 presents the future directions of this work and concludes the paper.
3. Requirements for supporting text
analysis computationally
The general situation in corpus-based text analysis
can be described as one where particular linguistic
phenomena observed in the corpus need to be
interpreted within a theoretical framework. This
interpretation may be an intellectual task typical
of scientific work for analysis, understanding and
explanation of texts. A tool for semi-automated
support analysis, then, needs to provide an integrated and user-friendly working environment whereby at least the following are supported:
• compiling and importing one or more text
corpora
• importing an already existing a – hierarchically or non-hierarchically – structured
knowledge classification schema
• definition and construction of one’s own
classification schema
• development or usage of one or more classification schemata for annotation and, thus,
knowledge categorization enabling annotation of text segments (words, groups of
words, etc.) according to the schema used
• interactive annotation by mechanisms that
are flexible and easy to use on text selections
• enabling work with more than one categorization schemata concurrently, so that the
analyst can work separately with different
levels of linguistic description, e.g. morphological, lexical, syntactic, etc., and at the
same time integrate information from different schemata (for most text analysis tasks
integration of different kinds of linguistic
knowledge is a requisite)
• the integration of already existing automatic
tagging/encoding tools for an initial annotation and enabling work with already annotated corpora, e.g. with SGML, PoS
annotation, etc.
• flexibly viewing already encoded text segments; this includes both selecting and arranging according to different criteria
(frequency of occurrences, encoded category
types, etc.) and presenting information with
readable layout styles (fonts, colours)
• concordance list presentations
• calculating different statistics for every word
and every annotated text segment and this on
the basis of the (hierarchical or non-hierarchical) relations within the categorization
schema, e.g. frequency of occurrence of
word tokens, word types, schema categories,
or demonstration of how these elements are
distributed in one or all texts
• providing solutions or mechanisms for exporting both annotated and non-annotated
information in different formats, e.g. SGML,
for further processing
• multilingual text analysis.
Current available tools and programs which may
support the above text analysis requirements can
be categorized according to corpus processing
tools, general text analysis tools with retrieval
capabilities, and annotation tools or programs, i.e.
taggers. To give just a flavour of these consider:
the DTTool for tagging corpora for anaphora
(Aone and Bennett, 1994), the INTEX corpus
processing system (Silberztein 1994), the COSMAS corpus storage, maintenance and access system with built-in corpora (al-Wadi, 1994), the
WordCruncher, a flexible PC-based text retrieval
corpus exploration tool (available from IKS,
Bonn), TACT (Bradley,1990) and Lexa (Hickey,
1992) for general text exploration with presenting
information in the form of concordance lists and
statistical tables, the knowledge extraction tools
KES and GRAAL (as described in (Ogonowski et
12
al. 1994)), the Xtract tool (Smadja and McKeown,
1990) for automatically extracting word collocational information and the WAG-coder (O’ Donnell, 1995) for systemic functional linguistics
oriented analysis.
The main weakness of the current systems is that
no support is offered in an integrated manner
where different analysis perspectives can be employed in one working environment and different
kinds of information extracted or made available
concurrently. Weaknesses usually include one or
more of the following: inflexible data model (e.g.
the data basis is the whole corpus and the corpus
parts such as texts, paragraphs, etc. are not separated), theory-dependence, domain-dependence,
applicable only for one level of linguistic analysis,
non-linguistically motivated, inability to either
import an already existing classification schema
or to provide support for consistently and efficiently building one, non-portability, inflexible
user interface, lack of multilingual support.
4. Text Analysis Tool with Object
Encoding
Based on the variant nature of the work and objectives described in section 2 we have designed
TATOE, a tool for a data-driven and multi-layered
text analysis. Our intention has been that the tool
is language and semantic domain independent,
that one or more categorization schema(ta) for
encoding can be flexibly developed by the user or
imported in the tool, and that the tool is portable
between different hardware platforms.
The implementation of TATOE took the requirements listed earlier as its starting point. We designed an object-oriented data model – using the
Smalltalk Frame Kit (SFK), an object oriented
modeling tool offering a spectrum of features to
make model descriptions operational (Fischer and
Rostek 1996) – for the representation of the different kinds of information to be stored in a kind of
‘personalized annotation database’. In such a database the text segments are linked to elements of
categorization schemata so that the database contents can be seen as a network of information units.
We use an object-oriented approach to handle a
fine granular network of highly interlinked information units (objects).
A general picture of the tool is given in figure 1.
Figure 1: A screen dump of TATOE
13
The variety of information presented on the screen
is arranged in five separate panes in order to keep
corpus texts and information about these text(s)
side by side. These panes are: the main text pane
in the middle, the list of texts at the top left, the
word list pane underneath it. On the top right there
is the list of categories above the list of marked
instances. With the exception of the text list, all
the lists contain different information with regards
to the words in the corpus, the categories used for
annotation and the already annotated instances
sorted out in lists.
The main characteristics of TATOE and its
functionalities are:
• Category set up and maintenance: The
categories to be used for the classification are
usually defined prior to the beginning of the
analysis, depending on the analysis objectives. It is possible to either import an already
available categorization schema or for the
analysts themselves to define a categorization schema. More than one schemata can be
defined and used concurrently. This enables
multi-layered text analysis (lexical, grammatical, etc.) A categorization schema in TATOE can be modified by adding, deleting,
renaming and merging categories. This holds
for schemata which have been developed
within TATOE as well as for those imported.
For our analysis tasks we have used Brill’s
rule-based PoS tagger (Brill, 1992) for the
English texts and imported the tagging categorization (together with the annotations) as
a separate categorization schema. We have
also imported the merged upper model (Henschel and Bateman, 1994) linguistic ontology as an additional classification schema.
• Searching and selecting: Searching for a
word is done by selecting the appropriate
item from a word list with all the corpus word
types. Searching for a text is performed by
selected the relevant text from the corpus text
list. Selection is performed either by choosing one of the items in the word list or by
selecting straight on the text. On text selection is also possible when viewing a concordance lists.
• Concordance lists: Together with a ‘classic’
concordance list there is also the possibility
for category concordance lists. Category
concordance lists include all the words which
have been already assigned to a schema category. Figure 2 shows such a concordance
list where all annotated instances for the category ‘artist’ are presented (in a black and
white graph the highlighted instances are
shown in white). Rather than using positioning to indicated the selected items in the
concordance list in TATOE colour is used.
The reason and advantage of this is the possibility to justify the left and right context of
the selected item ‘on the fly’.
Figure 2: Type concordance list of all instances belonging to a selected schema category
14
• Annotating: annotation is possible for either
a single occurrence of a selected item or all
its (identical) occurrences in the corpus.
Mark up of a selected item is possible either
from the full text display or from the concordance display. The marked up items are then
highlighted (in colour) so that they are distinguished from the rest of the text items
which have not been annotated yet.
• Qualitative and quantitative information:
upon selection of a marked up instance or a
schema category TATOE provides qualitative information, e.g. relations such as broader, part-of, etc. holding between instances
or between single schema categories. Selection of an item from the word list or the list
of all schema categories provides quantitative information, e.g. frequency of occurrence
information or total number of marked instances.
• Viewing: Performed, or already existing, annotation can be viewed either in the form of
a list of all annotated items or as a list of
occurrences in their context, i.e. KWIC lists,
or as statistical tables for the already annotated data (with each category together with the
total number of tokens or of types assigned
to it) or as a graph with all the relations
between the categories used represented in a
semantic network form. An example of the
latter is given in figure 3, where part of the
upper model categorization schema is shown
graphically.
Figure 3: Graphical presentation of categorization schema categories in TATOE
15
• Checking and correcting: In TATOE consistency is checked by selecting a text segment and viewing all its instances in a
concordance list. If the intention was that
they were all assigned the same category and
this has not happened, then the analyst can
mark up the remaining instances accordingly. Alternatively, if some of the text segments
should have been assigned a different category than the one they are shown to belong
to, these can be deselected and then selected
again for the category intended. Correction
follows the selection and deselection procedure, when the marking up is concerned.
• Frequency information: Statistical information is supplied in a number of ways: (i)
word frequency lists which are sorted alphabetically, or according to the highest frequency of all occurrences in the whole corpus or
to the number of texts or paragraphs one
word occurs (note that frequency of occurrence as shown in the word list of figure 1, is
sorted according to all occurrences of a word
in the whole corpus indicated by the far right
number, frequency of occurrence according
to the total number of paragraphs a word
occurs, indicated by the middle number, and
frequency of occurrence according to the total of corpus texts a word occurs, indicated
by the far left number), (ii) frequency list of
all the categories used for the encoding, (iii)
frequency tables of all schema categories (for
each schema categorization) together with
the total of the text segments encoded for
each one of them.
Technical characteristics
TATOE is implemented in the Smalltalk-based
programming environment VisualWorks (from
ParcPlace). It can be used on various hardware
platforms, e.g. Suns, Sparcstations, PCs and Macs.
Future work and conclusions
Currently TATOE is a prototype and its scaling up
for handling efficiently very large amounts of
textual data is one of the most important tasks. We
plan to test its usability for more applications
involving corpus-based text analysis. We expect
that new applications will result in refining the
TATOE’s data model to accommodate the specific needs of these applications.
Our intention is to develop a tool which can be
used for multilingual corpus-based text analysis
and we therefore plan to test its usability for a wide
range of languages apart from the ones it has
already been used for.
Regarding the functionalities of TATOE, we are
working on adding search capabilities of complex
formal patterns for extraction of collocational information in a more flexible way.
At present, we are investigating exporting the
annotated information according to Text Encoding Initiative guidelines. This has to take into
account that annotated text segments may often
overlap.
We have presented a tool for corpus-based text
analysis. Such analysis often requires to allow for
the possibility to view the data from different
perspectives. TATOE enables importing and
maintaining and developing different categorization schemata for annotation, and the utilization of
the different categorization structures for calculating a variety of cumulative statistical information.
References
Alexa, Melpomeni (1993): Corpus-based sublanguage analysis for a multilingual text generation system. Ph.D. Thesis, CCL, UMIST,
Manchester, 1993.
Alexa, Melina (1995): Making principled selections: a methodology for register analysis and
description for text generation. 22nd International Systemic-Functional Congress, Beijing,
China, July 1995.
al-Wadi, Doris (1994): COSMAS Benützerhandbuch. XII/2785. Institut für deutsche Sprache,
Mannheim, Germany, 1994.
Aone, Chinatsu and Scott W. Bennett (1994):
Discourse tagging tool and discourse tagged
multilingual corpora. In Proceedings of International Workshop on Sharable Natural
Language Resources (SNLR), Nara, Japan,
1994.
Bateman, John A. (1996): KPML: The KOMETPenman (Multilingual) Development Environment, Release 0.9. Technical report,
Institut für Integrierte Publikations- und Informationssysteme (IPSI), GMD, Darmstadt,
March 1996.
Bradley, John (1990): TACT: User’s Guide. Technical report, University of Toronto, 1990. Version 1.2.
Brill, Eric (1992): A simple rule-based part of
speech tagger. Proceedings of the Third Conference on Applied Natural Language Processing, 1992.
Fischer, Dietrich and Lothar Rostek. (1996): SFK:
a Smalltalk Frame Kit. Technical report,
GMD/Institut für Integrierte Publikations- und
Informationssysteme, Darmstadt, 1996.
Henschel, Renate and John Bateman (1994): The
merged upper model: a linguistic ontology for
German and English. Proceedings of 25th International Conference on Computational Linguistics (COLING 94), Kyoto, Japan, August
1994.
Hickey, Raymond (1992): Lexa, corpus proces16
sing software. Technical report. Report series:
the Norwegian Computing Centre for the Humanities, 1992. Vol. 57, 58, 59.
O’Donnell, Mick (1995): From corpus to coding:
semi-automating the acquisition of linguistic
features. Proceedings of the AAAI Spring
Symposium on Empirical Methods in Discourse Interpretation and Generation, California,
March 27–29 1995. Stanford University.
Ogonowski, Antoine, Marie Luce Herviou and
Eva Dauphin: Tools for extracting and structuring knowledge from texts. Proceedings of
the 15th International Conference on Computational Linguistics (COLING 94), vol. II, pp.
1049 – 1053, Kyoto, Japan, 1994.
Rostek, Lothar and Wiebke Möhr (1994): An editor’s workbench for an art history reference
work. Proceedings of the ACM European Conference on Hypermedia Technology, Edinburgh, ACM, New York, NY 1994.
Rostek, Lothar, Wiebke Moehr and Dietrich Fischer (1994): Weaving a Web: the structure
and creation of an object network representing
an electronic reference network. Proceedings
of Electronic Publishing (EP) ’94, pp. 495
–506, 1994. Special issue of the International
Journal of Electronic publishing – organization, dissemination and design, Volume 6(4).
Silberztein, Max D. (1994): Intex: a corpus processing system. Proceedings of the 15th International Conference on Computational
Linguistics (COLING 94), vol II, pp. 579 –
583, Kyoto, Japan, 1994.
Smadja, Frank A. and Kathleen M. McKeown
(1990): Automatically extracting and representing collocations for language generation.
Proceedings of the 28th Conference of the
Association for Computational Linguistics,
ACL-90, pp. 252 – 259, Pittsburgh, 1990.
Teich, Elke and John A. Bateman (1994): Towards an application of text generation in an
integrated publication system. Proceedings of
the Seventh International Workshop on Natural Language Generation, Kennebunkport,
Maine, USA, June 21–24, 1994.