JCMC 9 (1) November 2003
Collab-U  CMC Play  E-Commerce Symposium Net Law  InfoSpaces Usenet
  NetStudy VEs VOs O-Journ HigherEd Conversation Cyberspace Web Commerce
Vol. 6 No. 1  Vol. 6 No. 2  Vol. 6 No. 3  Vol. 6 No. 4  Vol. 7 No. 1  Vol. 7 No. 2  Vol. 7 No. 3  Vol. 7 No. 4  Vol. 8 No. 1  Vol. 8 No. 2

Bilingual Newsgroups in Catalonia: A Challenge for Machine Translation

Salvador Climent, Joaquim Moré, Antoni Oliver, Míriam Salvatierra, Imma Sànchez, Mariona Taulé*and Lluïsa Vallmanya
Internet Interdisciplinary Institute Universitat Oberta de Catalunya
* Department of Linguistics Universitat de Barcelona

Abstract

This paper presents a linguistic analysis of a corpus of messages written in Catalan and Spanish, which come from several informal newsgroups on the Universitat Oberta de Catalunya (Open University of Catalonia; henceforth, UOC) Virtual Campus. The surrounding environment is one of extensive bilingualism and contact between Spanish and Catalan. The study was carried out as part of the INTERLINGUA project conducted by the UOC's Internet Interdisciplinary Institute (IN3). Its main goal is to ascertain the linguistic characteristics of the e-mail register in the newsgroups in order to assess their implications for the creation of an online machine translation environment. The results shed empirical light on the relevance of characteristics of the e-mail register, the impact of language contact and interference, and their implications for the use of machine translation for CMC data in order to facilitate cross-linguistic communication on the Internet.

Introduction

Catalonia is an autonomous bilingual region of Spain1 where the native language, Catalan, is being displaced to a great extent by Spanish when Catalan speakers communicate with members of different linguistic groups. Catalan speakers tend to use Spanish if they know the addressee is not a native Catalan speaker, even if the addressee is able to understand Catalan. In addition, they tend to use Spanish as the default language when they are unsure of the addressee's level of understanding of Catalan. In computer-mediated communication (CMC), the tendency of Catalan speakers to use Spanish by default is even more noticeable. Thus, we reasoned that the gradual replacement of Catalan by Spanish could be prevented if those writing in Catalan felt that, thanks to human language technologies, they did not need to shift to Spanish. The existence of a machine translation system that could automatically translate e-mails into the language of the addressee would make it unnecessary for users to write in Spanish. Users' confidence in the translation quality would lead to their using their own language regardless of the other's language.

With this idea in mind, we and other colleagues founded the INTERLINGUA project. INTERLINGUA aims to promote machine translation (MT) to enable users to employ their native languages on the Internet. The project is not limited to Catalan and Spanish, but could be used for any pair of languages.2 Hence, INTERLINGUA is a project to facilitate communication between people who speak different languages.

However, MT systems are currently not capable of translating spontaneous e-mails accurately. Such systems work reasonably well if the input is in a standard form, but as we will see in this study, e-mails often contain spelling mistakes, typing errors and the like. Moreover, the systems are not prepared to cope with the creative and spontaneous use of language typical of e-mails, especially in a bilingual community. Thus, e-mail presents certain new challenges for MT. On the one hand, non-standard forms need to be converted into standard forms that can be recognized by the MT system, a challenge that the MT community has not yet addressed (Climent, Moré & Oliver, 2003). On the other hand, the project must also take into consideration the use of language in a bilingual society.

In this paper, we present the results of a study of e-mails written on the UOC Virtual Campus, carried out as part of the INTERLINGUA project. The e-mails were first translated from Catalan to Spanish or vice versa using an MT system. On the basis of the translation errors, we classify and quantify the linguistic features in the original messages that have a negative impact on translation quality, stressing those features that involve bilingualism, and analyze their distribution in the two languages.

The Sociolinguistic Situation in Catalonia

Catalan is still recovering from the extreme measures imposed under the rule of Spain's former dictator Francisco Franco, who declared Spanish the only official language and prohibited Catalan in the press, broadcasting, theatre, schools, street signs, advertising, shop signs, etc. Since the return of democracy after Franco's death, Catalan and Spanish have been co-official languages in Catalonia, and linguistic and educational policies have attempted to redress the situation of Spanish dominance over Catalan. However, according to official statistics, although these policies have improved the levels of people's understanding and ability to speak Catalan, they have not succeeded in improving levels of spontaneous usage. These levels are shown in Table 1. The data on people's understanding and ability to speak are taken from a survey by the Centro de Investigaciones Sociológicas (Sociological Research Centre; CIS, 1998). The usage data are taken from an analysis by Cerdà (2001), in the section entitled "Lengua predominante y competencia lingüística " (Predominant language and linguistic competence) from the same CIS survey.

Understand
Catalan
Speak
Catalan
Spontaneously
use Catalan
Spontaneously
use Spanish
Spontaneously
use both
97%
79%
41%
43%
16%

Table 1. Percentage of the population in Catalonia able to understand and speak Catalan and Spanish (Cerdà, 2001; CIS, 1998)


Another survey, carried out by the Institut d'Estadística de Catalunya (Catalonian Institute of Statistics) (IDESCAT, 2001), shows that the percentages for reading and writing Catalan are lower than those for oral comprehension and use. These are summarized in Table 2.

Understand
Catalan
Speak
Catalan
Read
Catalan
Write
Catalan
95%
73%
72%
46%

Table 2. Percentage of the population in Catalonia able to speak, read and write Catalan (IDESCAT, 2001)


The ability to read and write is obviously essential in CMC; thus, the growth of Spanish at the expense of Catalan is expected to be greater in this form of communication than in spoken contexts.

These statistics support the impression that Catalans themselves have, namely that the Catalan language is not essential for living in Catalonia. Moreover, the situation has worsened in the last decade due to massive immigration from Morocco, Sub-Saharan Africa and Latin America. According to the CIS survey, 18% of the immigrant population in Catalonia did not understand any Catalan at all in 1998.

Immigrants learn Spanish because it is the official language of Spain and a widespread means of communication. They do not feel compelled to learn Catalan due to the fact that, on the one hand, Spanish is omnipresent in mass media, books, shops, law, etc., and on the other hand, Catalans communicate with them in Spanish. The tendency of Catalan speakers to shift to Spanish when the addressee uses Spanish is an example of code-switching. The use of Spanish when addressing someone whose origin is unknown or even someone identified as a Spanish speaker, despite their being able to speak Catalan, is considered a sign of politeness. Many Catalan people also consider Spanish more prestigious, and view it as more important than Catalan because it is a language of wider communication. Some people also code-switch to Spanish because they feel more competent in that language (Pujolar, 2000).3

The presence of code-switching on the Internet is particularly apparent, as interactions between Catalans and those from outside Catalonia through chats, e-mail, etc. become increasingly common. Aside from the fact that Spanish has a far greater presence on the Internet than Catalan, the possibility of addressing a wider group of people from all over Spain and the world encourages the use of Spanish when writing to someone for the first time, or when replying to a message in Spanish, even in cases where the addressee understands Catalan.

The use of new technologies, where Catalan is conspicuous by its absence, has the greatest effect on young, urban dwellers (Castells & Díaz de la Isla, 2001). The young who, with the return of democracy, were expected to lead the way in making Catalan the normal, everyday language of communication, have instead tended to use Spanish. Thus, communication over the Internet in Catalan is increasingly unlikely, in that the main group using this technology is urban youth. This trend goes against the established thought of many scholars who praise the Internet's ability to promote minority languages and multilingualism (Mehsching, 2000; Warschauer, 2000).

Language Choice in a Preliminary E-mail Sample

In order to assess the current status of Catalan on the Internet, we analyzed a sample of e-mails from the UOC-Catalonia4 Virtual Campus. Catalan is the institutionalized language of the Virtual Campus. That is, Catalan is the language used in educational material as well as by teachers when addressing students in the virtual environments. Although there are no official restrictions on the spontaneous use of other languages, it is assumed that the people who register for instruction through UOC-Catalonia will be fully competent in Catalan.

As a test bed for the research, several so-called Fòrums d'Informàtica (computer science newsgroups) were chosen. In these informal newsgroups, students exchange information and opinions related to computers, software, bugs, cheats, academic subjects, etc. These newsgroups are not, by themselves, representative of text-based, asynchronous computer-mediated communication in Catalonia as a whole. Both linguistically and sociolinguistically, the communicative situation in Catalonia is too diverse and complex to be represented by a single group. Nonetheless, we believe that these newsgroups are good examples of contemporary practice, and that by analyzing them we can learn many things about how languages are currently used in Catalonia in online communication.

Although the official language of UOC-Catalonia is Catalan, messages and replies are posted in the forums in both Catalan and Spanish, sometimes mixing the two languages. In an attempt to quantify the degree of code-switching between messages (as opposed to within messages), we analyzed all messages mailed to the forums between July and December 2002: 533 messages sent by 254 users (an average of 2.1 messages per user). In this sample, 76% of the messages were in Catalan and 24% were in Spanish. To infer the degree of code-switching in the group, we considered only the 189 e-mails that were replies made by those users defined as spontaneous in one or the other language. The criteria for defining spontaneous users are given in Figure 1. We considered those users who wrote initiating e-mails (i.e., not replies to other e-mails) in both languages indistinctly as indifferent. Finally, we considered those users who replied to e-mails originally written in A in the same language, but who did not write any initiating e-mails, as undetermined.

Users (U) are regarded as spontaneous in language A, if:
  • They wrote only in A and not all of their e-mails were replies to initiating e-mails
    written in A (no code-switching), and
  • they replied in A to e-mails written in B; or
  • they generally wrote in A, although they replied in B to certain e-mails written in B.

Figure 1. Definition of a “spontaneous user” of a particular language


Table 3 shows the results of applying this classification scheme to the messages in the sample.

Spontaneous Catalan Users
Spontaneous Spanish Users
Indifferent
Undetermined
68.9%
18.1%
1.2%
11.8%
Reply to e-mails
written in Spanish
Reply to e-mails
written in Catalan
In Catalan
In Spanish
In Catalan
In Spanish
57.1%
42.9%
15.4%
84.6%

Table 3. Classification of messages in terms of spontaneous language use


Although these results may not be statistically significant because the sample size is small, they suggest that code-switching between messages is an important phenomenon among spontaneous users of Catalan (43% of messages), but less so among spontaneous users of Spanish (15% of messages). This, and the fact that only 69% of messages were sent by spontaneous Catalan users, appears paradoxical when we recall that, in the environment studied, Catalan should be the only language used.

In other UOC-Catalonia environments, the dominance of Spanish over Catalan is even more pronounced. Many users of the Ph.D. virtual classrooms are students who do not live in Catalonia (but rather come from other parts of Spain or South America), so they need not be competent in Catalan. As a result, activities in 14 of the 15 Ph.D. classrooms are carried out exclusively in Spanish, despite the fact that structural and institutional information is given in Catalan.

All of this evidence suggests a trend that is endangering Catalan as a language for communication on the Internet.

The Role of Machine Translation

MT systems allow users to employ their own language, regardless of whether the addressee can read it. If the quality of the translation is sufficiently high, the sender can trust the MT system and avoid having to code-switch in order to guarantee communication. In this way, the better the quality of the translation, the more likely the sender is to use his or her native language.

However, to guarantee quality in the translation of e-mails, certain factors need to be taken into account. The MT systems currently in operation depend on the input text being correct and standardized; i.e., the systems' rules and lexical databases are only able to recognize standard words and correctly written texts. Even when working with standardized texts, machine translation systems make mistakes; the greater the structural differences in the languages involved, the more mistakes the systems make. Thus it is generally assumed that MT systems are not able to produce perfectly correct translations, merely approximate translations that allow the addressee to understand the gist of the text. In the case of non-standardized texts, such as e-mails, the quality is expected to be even lower. Moreover, messages written by bilingual users may contain further deviations: Such users might mix languages when quoting or linking to previous articles, employ words of the other language, make spelling and grammatical mistakes due to language interference, and so on.

Currently, MT specialists take for granted that any text to be submitted for automatic translation should be manually pre-edited to overcome errors and deviations from the standards, as well as post-edited to correct the remaining mistakes. However, in e-mail communication, human pre- and post-editing are not feasible, as the system has to work in real time and completely automatically.

For this reason, an in-depth analysis of the e-mail register is needed to shed light on the specific problems an MT system might be required to overcome in order to produce a good quality translation in an unsupervised environment (no pre- or post-editing).

The E-mail Register

Researchers in the field of CMC have studied various aspects of the e-mail register. We will focus here on the relationship between the linguistic structure of e-mails and non-standard features.5

According to Herring (2001), most non-standard features in English e-mails are deliberate choices made by users to: Murray (2000) claims that CMC in general uses what she calls "simplified registers," characterized by (among other features) short sentences, special lexicon and feedback devices that facilitate the reader's comprehension, as well as simplifications that may include the use of abbreviations and the omission of articles, pronouns, and copulae. According to Murray, the technology constrains time and space. CMC relies on typing, computer, and network speed, and CMC provides no visual paralinguistic or nonverbal cues. Consequently, CMC users employ strategies that reduce the time needed to write messages or that substitute for the lack of paralinguistic and nonverbal cues.

According to Yates and Orlikowski (1993), the mimicking of spoken language features in e-mail results in unconventional orthography, such as textual indications of emphasis (e.g., If an implementation DOES support vectors...), informal words typically used in speech (e.g., groove, stuff) or syntactic informality often taking the form of incomplete sentences and conversational cadences, usually combined with word choice and punctuation in order to simulate oral communication, as in Hmm, I see...

As for the creative use of language, Alonso, Folguerà and Tebé (2000), focusing on the Catalan lexicon used on the Internet ("Internet slang"), identify a category of informal, expressive lexical elements, such as correu tortuga ("snail mail;" literally "tortoise mail") and emili ("e-mail" referred to humorously, due to its resemblance to the proper noun Emile). These lexical items are common in Internet lexicon and convey a recreational, creative, ironic or informal attitude.

For Fais and Ogura (2001), there are features that are exclusive to e-mail,6 causing it to differ significantly from both formal textual and spoken language. These include visual and discourse-level phenomena such as the following:

  1. A highly idiosyncratic use of indentation and spacing to mark paragraph shifts, in such a way that a difference in paragraphing is typically interpreted as a cue to a different topic.
  2. Openings and closings: "Closings are typically formalized and devoid of meaning. Openings, on the other hand, contain information about the addressee(s). (...) The variability in format for openings and closings also makes their recognition a difficult problem."
  3. Use of visual strategies to capture aspects of spoken utterances, such as:
The aforementioned studies tend to focus on new, intentionally expressive devices. This may be because, as Herring (2001) concludes,
[a]ctually, although computer-mediated language often contains non-standard features, only a relatively small percentage of such features appears to be errors caused by inattention or lack of knowledge of the standard language forms. (p.616)
However, in the e-mail register, texts may contain many unintentional language mistakes if the user writes quickly and carelessly. Likewise, the assertion that lack of knowledge of standard language forms is relatively minor may be true in monolingual English-speaking environments; however, in a bilingual community, lack of knowledge of standard forms of one language or the other can be significant. For instance, in Catalonia, some users writing in Catalan have less of a command of Catalan than Spanish, in that they may not have studied Catalan at school or may not be used to reading Catalan. Their lack of knowledge may be displayed, unintentionally, in e-mails. Another factor that has been inadequately discussed in the CMC literature is interference between two languages in the messages of bilinguals.

Research Questions and Methodology

The study was motivated by the following questions:
  1. Which non-standard linguistic features are responsible for bad translations when using an MT system?
  2. To what extent are such features intentional?
  3. To what extent are they related to lack of knowledge of standard language norms?
  4. Do users in our study write in Catalan as well as they do in Spanish?
  5. What is the influence of language interference on the features that cause translation problems?
To address these questions, we proceeded according to the following steps:
  1. Evaluation of the performance of an MT system (identification of translation problems)
  2. Linguistic classification of non-standard features that cause translation problems
  3. Quantification of the frequency and distribution of the problem features in the corpus
  4. Interpretation of the results
These procedures were applied to a corpus containing messages in both Catalan and Spanish: The Catalan messages were translated into Spanish, and the Spanish messages were translated into Catalan for the purposes of the evaluation and analysis.

Evaluating Translation Quality: The Impact of Non-Standard Features

The MT system used in the INTERLINGUA project is Sail-Labs Incyta ES/CA, an application to translate between Catalan and Spanish and vice versa which is based on the prestigious METAL system developed by Siemens. This system was evaluated in the first stage of the study in order to find out how it worked and what its shortcomings were. By analyzing the results of the evaluation we would know to what extent non-standard features were responsible for poor quality translations, and to what extent these features were attributable to specific aspects of the e-mail-register or, alternatively, to the user's lack of awareness of standard language forms.

The evaluation of the system followed the ISLE international standards for MT evaluation (ISLE, 2000) and consisted of two processes: macro- and micro-evaluation (Van Slype, 1979).The macro-evaluation provided information about the acceptability of the translation system from a global perspective (intelligibility, fidelity, readibility of the e-mails translated). The goal of the macro-evaluation was to validate the translations performed in order to assess the usability of the system, without detailing its limitations. In contrast, the micro-evaluation identified the system's limitations and was required in order to establish a strategy for improvements. The micro-evaluation provided information about the origin of translation errors, whether they came from the system's shortcomings or from the user's writing. In this paper, we focus on the micro-evaluation.

In as much as the system translated segment by segment (roughly sentence by sentence), the micro-evaluation was carried out on text segments. The corpus prepared for the micro-evaluation amounted to 1239 segments in Catalan and 1128 segments in Spanish, taken from 129 randomly-selected e-mails for each language.7 The number of words amounted to a total of around 25,000. All segments were sent to the MT system so that the corpus was constituted by juxtaposing each segment with its automatic translation. The source of the corpus was the postings between July and December 2002 in four computer science newsgroups, where students ask for assistance, offer advice, announce events and so on. The reason for choosing these newsgroups is that they are the most active and provide the largest corpus size.

We developed a tool to perform the micro-evaluation, according to which the evaluation of each segment was carried out in five steps. First, a human evaluator had to judge whether the translation was intelligible or not without reference to the source segment. Second, given both the source segment and the translation, the evaluator had to decide whether the translation was faithful to the original in content, intelligibility and style. Third, if the translation was not fully intelligible or faithful, the evaluator had to grade the translation errors that led to the problem. We established four levels of translation error, based on Green's Rating Scale (Green, 1977): 1) a minor error (an error that affects the style), 2) an error that does not impair comprehension of the segment, 3) an error that leads to ambiguity, 4) serious errors (errors that make translation unintelligible). The fourth step was to analyze the original and the translation and to classify the translation error as either caused by the writer of the input, or by inadequate functioning of the system. If caused by the writer, the evaluator had to state whether the writer had expressed him- or herself infelicitously, written a syntactically incorrect sentence, or used a non-standard language form (a typing error, a spelling error, or an intentional or unintentional lexical deviation). The evaluator also had to consider "language interference," a category that affects expression, syntax and lexicon. If the translation error was caused by the system, the evaluator had to state whether the translation error was morphological or syntactical, or whether there were words, terms or expressions that were not translated or translated badly. Fifth, after having performed these steps, the evaluator could write comments that would be an important source of information for future improvements and research into e-mail writing and MT.

The evaluation was carried out by six bilingual language consultants employed full-time by the university, whose job it is to edit documents published at UOC for grammar and style. At least two evaluators examined each segment of text. The results were then collected and analyzed.

Data Analysis

In this paper, we only show the results concerning what the evaluators regarded as translation errors caused by the input. The errors caused by malfunctioning of the MT system are not presented here.

The analysis of the data was carried out by systematizing the classification and comments of the evaluators regarding the translation errors, and quantifying the results. Our aim was to determine which characteristics of the text were attributable to the writer's intention to use language that differed from formal norms, and which characteristics were attributable to other factors, focusing on language contact, a significant aspect of the area under study. Another important aim was to quantify each type and subtype of phenomenon. We believe this approach to be innovative, as the CMC literature to date has often pointed out certain phenomena without quantifying their actual relevance. We think that certain phenomena that may have been overvalued because of their novelty actually have little effect, while, on the contrary, other important phenomena have been neglected.

Due to the size of the corpus studied, the background of the users, and the specific nature of their communication needs, the conclusions we have drawn from the analysis of the results cannot describe e-mail communication in general. However, as will be seen below, we can infer certain interesting points.

Classification Scheme

We have empirically classified the linguistic characteristics that cause translation errors into three broad areas: (1) unintentional non-standard features, (2) intentional non-standard features and (3) terminology.

The full classification scheme used for the empirical analysis of the corpus is summarized below. Following that, we describe each category and subcategory in turn.

1. Unintentional non-standard features
1.1 Mistyping
1.2 Deviations from prescriptive language norms
         1.2.1 Orthographic
                   1.2.1.1 Accents
                   1.2.1.2 Phoneme-grapheme confusion
                   1.2.1.3 Composition and separation symbols
                   1.2.1.4 Capitalization
                   1.2.1.5 Errors in abbreviations and acronyms
         1.2.2 Lexical
                   1.2.2.1 Barbarisms
                   1.2.2.2 Recurrent mix-ups
                   1.2.2.3 Oral reproduction
                   1.2.2.4 Loan-word errors
         1.2.3 Syntactic
         1.2.4 Cohesion
                   1.2.4.1 Verb tense errors
                   1.2.4.2 Anaphoric errors
                   1.2.4.3 Punctuation errors
2. Intentional non-standard features
         2.1 Language shift
                   2.1.1 Lexical
                             2.1.1.1 Expressive
                             2.1.1.2 Terminological
                   2.1.2 Phrasal
         2.2 New forms of textual expressivity (characteristic of the e-mail register)
                   2.2.1 Orthographic
                             2.2.1.1 Orthographic innovations
                             2.2.1.2 Systematic lack of accentuation
                   2.2.2 Lexical
                             2.2.2.1 Internet user vocabulary
                             2.2.2.2 Informal (oral-like) language
                             2.2.2.3 Prosodic reproduction
                             2.2.2.4 Shortenings
                   2.2.3 Visual
                   2.2.4 Pragmatic
                   2.2.5 Simplified punctuation
                   2.2.6 Simplified syntax
3. Terminology
         3.1 Domain terminology
         3.2 User community terminology

Main categories: Unintentional non-standard features (1), intentional non-standard features (2) and terminology (3)

Our first main category is unintentional non-standard features (1). These differ from intentional non-standard features (2) in that they are not deliberately chosen by the writer. In some cases, doubt arises as to whether a non-standard feature is intentional or not. These cases are considered in relation to the context of the e-mail as a whole. If the case appears to be embedded in a system of coherent odd features, it is classified not as unintentional but as intentional (2). For instance, with regard to accents, if only one or a few words in the e-mail lack the necessary accentuation, we classify it as unintentional. However, when the user does not use accents at all in the message, or where the lack of accentuation is consistent with a rationale, then we regard this as a kind of voluntary deviation, i.e., systematic lack of accentuation (2.2.1.2).

Accordingly, our second main category, intentional non-standard features (2), is characterized by deliberate choice. One main group of intentional deviations is that which, according to the literature, defines the e-mail register itself: new forms of expressivity (2.2)-oral patterns, shortenings, simplified punctuation or syntax, specific pragmatic resources, visual information, etc. The other main group is language shift (2.1), i.e., the use of words or constructions from other languages, even though well-known equivalents exist in the language in which the text is written. Both the categories and their subcategories are explained below.

Last, we classified as terminology (3) vocabulary which is specific to either the domain of knowledge and communication (in this case, computer science) or the particular user community under consideration (in this case, UOC students). This differs from vocabulary that can be considered part of the general register of Internet users, classified as (2.2.2.1) under new forms of expressivity.

Mistyping (1.1)

These are mainly deviations caused by neighboring key strikes (*Cstalonia instead of Catalonia), extra strikes (*Caatalonia), inverted strikes (*Catlaonia), missing strikes in a word (*Ctalonia), or connecting two words *toCatalonia). We also include here the mistyping of a symbol that is similar to one that the user intends to strike; a typical example is the use of accents instead of apostrophes.

Deviations from prescriptive language norms (1.2)

In this case, the deviation is caused by users not being aware of a rule or a norm of the language they are using. This occurs at different linguistic levels, as described below. In Catalonia, a number of these deviations can be seen to be caused by language interference, although it is difficult to say exactly how many, since it depends greatly on the social and educational backgrounds of each individual. We return to this point later.

Orthographic deviations (1.2.1)

We have found different types of orthographic deviations, from erroneous capitalization (i.e., asystematic non-capitalization) to errors in writing acronyms and abbreviations or in the use of certain characters (e.g., apostrophes and hyphens in Catalan to affix clitics, as in *dona'me-l for dona-me'l, "give + to me + it"). However, the most common orthographic deviation comes from accentuation and phoneme-grapheme confusions (*andavant; *adreçes; *trovar instead of endavant, adreces, trobar), typically when one phoneme can be spelled by many graphemes. For instance, both 'a' and 'e' can represent the schwa sound in Catalan, and writers sometimes choose the wrong letter, as in the case of *andavant (forward), which should be written endavant. Similarly, 's,' 'c' and 'ç' can all represent the /s/ sound, and writers sometimes choose the wrong option, e.g., *adreçes should be adreces (addresses).

This happens in our corpus in four cases: (i) a/e and o/u alternation to represent the schwa sound and the unstressed /u/ sound; (ii) c/s/ç to spell /s/; (iii) b/v for /b/; and (iv) confusion in the use of digraphs: s/ss to spell /s/, and l/l·l and n/nn. The digraphs 'l·l' and 'nn' represent a combination of two /l/ and two /n/ respectively, which are prescribed by spelling norms but are hardly ever pronounced in the oral language. For instance, speakers pronounce a single /l/ when saying 'pel·licula' (film).

Lexical deviations (1.2.2)

We have found four types of lexical units that differ from standard language use. The first are what in Catalonia are called barbarisms (1.2.2.1), words or lexical constructions that the speaker believes are genuinely Catalan but which in fact are Spanish. Examples of these are *insertar (instead of inserir, "to insert") and *recent (instead of acabat de fer, "fresh"). These are archetypal cases of interference between languages in contact. In Catalonia, they also occur in Spanish due to the influence of Catalan, as in *antes de nada instead of en primer lugar ("first of all"). This particular example is caused by the speaker translating the lexical construction word for word from the Catalan equivalent abans de res (abans = antes = "before," de = de = "of," res = nada = "nothing").

The second type is lexical mix-ups (1.2.2.2) which are caused by similarity of form but difference of meaning, e.g., si no/sinó ("but"/"otherwise"), per què/perquè ("why"/"because"), per/per a ("for" or "by"/"in order to") in Catalan, and a parte/aparte ("in part"/"apart") in Spanish. These are very common in some people's writing. Some of these mix-ups may also be caused by language interference due to false analogies between similar forms in Spanish and Catalan.

The third type of lexical deviation is caused by attempts to reproduce oral forms in writing (1.2.2.3). There are different subtypes, but all of them are distinct from phoneme-grapheme confusion (1.2.1.2), in which one phoneme can be spelled by two or three alternative letters, thus constraining the deviation in terms of available options. The scope is wider in the case of oral spelling, in as much as it might affect several phonemes/graphemes, or the whole word, thus changing the overall form of the lexical unit; for this reason it has been classified as 'lexical.' Typical cases are vols or a veure, Catalan words which some speakers pronounce /bos/ and /abere/, so that those speakers sometimes mistakenly write them as *vos and *avere. Another case is the pronunciation of donés /dunes/ (a subjunctive form of "to give") with an epenthetic velar consonant, /dunges/, thus leading the word to be written as *dongués. An example in Spanish is *osea instead of o sea; in this case the oral reproduction consists of converting two words into one, thus reproducing the seeming lack of spacing between words in continuous speech. Deviations caused by oral reproduction are dialect-dependent, in as much as pronunciation in different dialects resembles to a greater or lesser extent the standard in Catalan or Spanish.

Last, we also found cases of errors in the spelling of loan words (1.2.2.4), e.g., mistakenly writing the English word cookies as *cookis, or Access (the database software) as *Acces.

Syntactic deviations (1.2.3)

This category covers the non-standard use of grammatical categories (e.g., incorrect choice of verbal mood, as in the use of infinitive instead of imperative in *decirme instead of decidme ["tell me"] in Spanish) and other cases of syntactic ill-formedness. Relevant examples of the latter are the omission or addition of pronouns, prepositions or other function words, as in Catalan's *jo vull ("I + want") instead of jo en vull ("I" + direct object pronoun + "want") to mean "I want (that thing)" or, in Spanish, *pienso de ir for pienso ir ("I think I'll go"); as well as typical cases of lack of agreement (subject-verb, determiner-noun).

Although it is difficult to generalize due to the sparseness of this type of data in the corpus, it is clear that at least some syntactic deviations are caused by language interference, as in the first two examples above, where (i) the incorrect omission of the pronoun en in Catalan reflects Spanish norms, and (ii) the incorrect addition of the preposition de in Spanish reflects Catalan norms.

Cohesion deviations (1.2.4)

Textual cohesion is negatively affected in our corpus by inappropriate use of punctuation marks (colons, semicolons, hyphens, etc.), incorrect choices of verbal tenses to express temporal relations, and lack of concordance between pronouns and their antecedents.

Language shift (2.1)

As mentioned above, intentional deviations from language standards are classified into two main categories, the first being language shift, i.e., the voluntary use of words or phrases from other languages. In Catalonia, where there is close contact between Catalan and Spanish, language shift is very common in informal speech since all speakers have a degree of knowledge of both languages. A consequence is that sometimes, when speaking in language A, a lexical choice corresponding to language B comes naturally to the speaker's mind. Since the language shift does not usually affect communication, in as much as the interlocutor is also bilingual, the speaker uses the other language's word, or even sometimes a phrase, not by mistake, but for the sake of fluency or other expressive reasons. For instance, it is typical to swear in Catalan using Spanish jo or joder ("fuck"), or to say goodbye in Spanish using Catalan adéu.

This also happens to Catalan and Spanish speakers with third languages, especially English. Sometimes people say goodbye by using Italian ciao, express gratitude by using French merci, or ask for aid with English help.

Not every intentional use of language shift is expressive: Many shifts involve terminology which has either been learned or is better established in another language. A typical example is the use of English software instead of Catalan programari. This case is debatable, in as much as some speakers might simply be unaware of the existence of the Catalan term. However, we have classified these cases as intentional since we assume that our users either know the terminology of their field in their language (but that they still prefer using English terms), or else are aware that there must exist a word in their language for the term (but they do not wish to stop to think about it or look it up in a dictionary when they write e-mails). Finally, those foreign terms that lack a well-known equivalent in Spanish or Catalan have been classified as domain terminology (3.1).

Two characteristics should be highlighted about such language shifts in e-mails: (i) they are related to the written reproduction of informal speech, and (ii) they are related to language interference.

New forms of textual expressivity (2.2)

These are the features that, according to the literature, best define the e-mail register. We find here simple categories, such as visual resources (2.2.3)—typically, smileys; the pragmatic resource of dialogue simulation in quoting part of a previous message (2.2.4); and simplified punctuation (2.2.5) or syntax (2.2.6). We classified as simplified syntax cases involving the lack of a function word in intentionally telegraphic constructions, e.g., the lack of an article in M'adreço a aquest fòrum amb l'esperança de trobar tècnic disposat a... (instead of ...l'esperança de trobar un tècnic...) ("...I'm hoping to find [a] technician..."). To distinguish between punctuation errors (1.2.4.3) and simplified punctuation, we counted as the latter any lack of (expected) punctuation marks in e-mails lacking any punctuation at all.

Similarly, for accentuation, we counted those e-mails which did not have any accents at all separately, so that any lack of an accent within them was classified as a case of systematic lack of accentuation (2.2.1.2). Otherwise, when occurring in e-mails that did have accents, lack of an accent was counted as an error (1.2.1.1).

Systematic lack of accentuation is one subtype of "new orthography." The other main class (2.2.1.1) includes a wide range of innovations such as capitalization or the use of a range of symbols to show emphasis ( necessito ajuda URGENT ... "I need help URGENTLY; "no funciona!!??!! "it doesn't work!!??!!"), use of symbols as meaning components in words (tod@s covering both masculine and feminine genders instead of "todos y todas "), or the use of ['s] to pluralize acronyms, as in CD's.

The other main class under "new forms of expressivity" includes a variety of lexical units which are not found in formal texts (2.2.2). First, we have colloquial Internet user vocabulary (2.2.2.1) such as online, hoax, nick, àlies ("nickname") or xat ("chat"). These are usually English terms or adaptations from English. We have not classified English terms as language shifts or terminology as they clearly belong to an emerging Internet register more than to the specific domain of computer science.

The second subtype includes general-purpose informal vocabulary (2.2.2.2), typically used in speech but never in formal texts, e.g., mates ("maths") for matemáticas ("mathematics"), profe for profesor ("teacher, lecturer") or yuyu (a colloquial term in Catalan and Spanish for either feeling under the weather or unusual behavior).

Another class is that of intentional reproduction of spoken prosody (2.2.2.3) used as an expressive resource. For instance, modessssno contains a graphical reproduction of a very long [s] sound; this "word," which represents moderno ("modern/fashionable"), means something or someone pretending to be fashionable but who in fact seems ridiculous. We also include here reproduction of oral sounds such as hmmm (expressing doubt) or psé (indifference).

The last category is that of SMS message-like shortenings (2.2.2.4), such as tb instead of també ("as well") or k for que ("who, what, which...").

Terminology (3)

As expected, we found many examples in our corpus of terminology. This finding is crucial for machine translation, since these are words that are usually missing in the lexical databases of MT systems, and could therefore cause errors in translation. Most of the time, such terminology is associated with a specialized knowledge domain: in the case of our newsgroups, computer science. Thus we find words such as XML, disc dur ("hard disk"), and script. However, we also find terms particular to the community of users, UOC students. These include PACs (a kind of academic assignment) and MIC (an acronym for an academic subject, Multimèdia i comunicació, "Multimedia and communication").

Such terms cannot be considered characteristic of the CMC register: CMC features are expected to be found in any kind of newsgroup; however, in a newsgroup devoted to medicine, for example, we will find medical terms instead of terms for computer science. Furthermore, in a newsgroup devoted to computer science in another kind of community, e.g., professionals instead of students, we would not expect to find student vocabulary such as MIC or PAC.

Quantification of Features that Cause Translation Problems

Having classified the errors and deviations from standard language use found in the corpus, we present the quantitative results of the classification in Table 4. The columns labeled AF (absolute frequency) show the total number of occurrences of each category in the corpus. RF (relative frequency) shows the number of occurrences of each category per thousand words in the corpus. IT (impact on translation) indicates the high (H), medium (M) or low (L) expected impact of the category on the quality of translation, independent of the number of occurrences. The IT ratings were produced according to the level of translation error caused by each category. This level was established by the evaluators during the evaluation process, when they had to grade the translation error according to Green's Rating Scale.

CATALAN
SPANISH
IT
AF
RF
AF
RF
1. Unintentional non-standard features
512
46.7
322
30.7

1.1 Mistyping
92
8.4
55
5.2
H

1.2 Deviations from prescriptive language norms
420
38.3
267
25.4

1.2.1 Orthographic
296
27.0
169
16.1

1.2.1.1 Accents
233
21.2
149
14.2
H

1.2.1.2 Phoneme-grapheme confusion

49
4.5
2
0.2
H

1.2.1.3 Composition and separation symbols
3
0.3
0
0.0
H

1.2.1.4 Capitalization
9
0.8
7
0.7
L

1.2.1.5 Errors in abbreviations and acronyms
2
0.2
11
1.0
L

1.2.2 Lexical
54
4.9
19
1.8

1.2.2.1 Barbarisms
17
1.5
8
0.7
H

1.2.2.2 Recurrent mix-ups
5
0.4
4
0.4
H

1.2.2.3 Oral reproduction
29
2.6
7
0.7
H

1.2.2.4 Loan-word errors
3
0.3
0
0.0
M

1.2.3 Syntactic
36
3.3
48
4.6
H

1.2.4 Cohesion
34
3.1
31
2.9

1.2.4.1 Verb tense errors
8
0.7
3
0.3
M

1.2.4.2 Anaphoric errors
1
0.1
9
0.8
H

1.2.4.3 Punctuation errors
25
2.3
19
1.8
H
2. Intentional non-standard features
155
14.1
346
32.9

2.1 Language shift
24
2.2
46
4.4

2.1.1 Lexical
24
2.2
45
4.3

2.1.1.1 Expressive
5
0.4
4
0.4
M

2.1.1.2 Terminological
19
1.7
41
3.9
L

2.1.2 Phrasal
0
0.0
1
0.1
M

2.2 New forms of textual expressivity

131
11.9
300
28.6

2.2.1 Orthographic
71
6.5
250
23.8

2.2.1.1 Orthographic innovations
53
4.8
86
8.2
M

2.2.1.2 Systematic lack of accentuation
18
1.6
164
15.6
H

2.2.2 Lexical
36
3.3
39
3.7

2.2.2.1 Internet user vocabulary
8
0.7
18
1.7
L

2.2.2.2 Informal (oral-like) language
9
0.8
8
0.8
M

2.2.2.3 Prosodic reproduction
6
0.5
5
0.5
H

2.2.2.4 Shortenings
13
1.2
8
0.7
M

2.2.3 Visual
9
0.8
3
0.3
L

2.2.4 Pragmatic

2
0.2
3
0.3
L

2.2.5 Simplified punctuation
2
0.2
0
0.0
H

2.2.6 Simplified syntax
11
1.0
5
0.5
H
3. Terminology
396
36.1
437
41.6

3.1 Domain terminology
268
24.4
293
27.0
L

3.2 User community terminology
128
11.7
144
13.7
M
TOTAL
1063
96.8
1105
105.2

Table 4. Frequency and distribution of features that caused translation errors

Discussion

It appears that the e-mails in the sample are not only characterized by new forms of expressivity, as is often claimed in English CMC research, but also by at least as many unintentional infelicities, mistypings and deviations from prescriptive norms. Most of these deviations seem to be due to a weak awareness of the language, especially Catalan, as the data show that the number of deviations from prescriptive norms is noticeably higher in this language than in Spanish. We have analyzed only a specific group of highly-educated users, university students. It is likely that among less well educated users, the incidence of unintentional infelicities would be higher.

Analogy with Spanish sheds light on a number of deviations from the norms in Catalan, such as accentuation of the common ending -ia (e.g., *enginyería instead of enginyeria, "engineering") or phoneme-grapheme confusions such as the failure to use 'ss' to represent the phoneme /s/, as in *asociació instead of associació, "association." The interference is clear in such cases, since the norms of Spanish demand both accentuation of '-ía' and use of 's' instead of 'ss.' Language interference is also evident in the lexicon (barbarisms) and explains certain recurrent confusions. In Spanish, language interferences are mainly cases of analogy in accentuation as well (Catalan: exàmens ("exams")-Spanish: *exámen instead of examen). However, the impact of these spelling confusions is not as great as it is in Catalan. The number of barbarisms in Spanish e-mails is also lower, and in recurrent mix-ups the difference disappears.

Many types of deviations cannot be explained in terms of interference. However, language interference is the single most important influence on the deviations found in the corpus. We would hypothesize that each of the following types of deviation was caused mostly or entirely by linguistic interference: incorrect accentuation, grapheme-phoneme confusions, barbarisms, recurrent mix-ups, syntactic errors and language-shift deviations. This hypothesis is supported by the comments of the language experts who evaluated the data, all of whom registered statements to this effect. Adding up these categories results in an estimate that 49.1% of the errors in Catalan and 31.3% of the errors in Spanish were caused by language interference. If we simply focus on deviations from norms (not counting language shifts), we find that as many as 59.4% of deviations from prescriptive norms in Catalan and 50.6% in Spanish were plausibly caused by interference. The high incidence of interference-induced errors and deviations is no doubt due to the fact that the users in our sample must deal with language contact on a daily basis. This situation represents a unique challenge for the application of MT to e-mail communication in Catalonia.

On the other hand, the ratio of intentional language shifts is not very significant (only 2.2 per thousand words in the corpus). Therefore, it appears that the main influence of linguistic interference in spontaneous texts is that it causes deviations from prescriptive norms.

Adaptation to intentional deviations through customization of the MT system in both directions would be worthwhile as well. Among new forms of expressivity, the feature which best characterizes the register is new orthography. New orthography is much more noticeable in Spanish (83.3% of intentional features) than in Catalan (54.1%). Another interesting point is that there seems to be greater impact in terms of intentionality and the creative use of language in Spanish than in Catalan. In Spanish, fully intentional non-accentuation is more evident. Likewise, there are more orthographic innovations in Spanish than in Catalan.

Lexical forms of new expressivity, considered together, have some effect in characterizing the register, although their relative frequency in terms of the corpus as a whole is low (3.3 for Catalan and 3.7 for Spanish). In contrast, visual and pragmatic resources, simplified syntax and simplified punctuation, despite having been paid a great deal of attention in the CMC literature (Herring, 2001; Murray, 2000), appear to be scarcely significant in either direction.

Another aspect often characterized as significant is oral patterns. However, the impact of the features related to this in our study did not reach expected levels. The features selected as oral patterns were barbarisms, oral reproduction, language shift, informal oral-like language and prosodic reproduction. It is not completely clear whether all of the barbarisms or language shift reproduce oral behavior, but they are included here as they refer to vocabulary that is used when speaking, but not usually when writing a formal text. Counting all of these features as oral patterns, they represent 12.7%, in Catalan, and 11.0%, in Spanish, of all of the non-standard features. If unintentional features are set aside and we concentrate on intentional aspects, oral patterns are, in Catalan, 25.1%, and, in Spanish, 17.0% of all of the intentional deviations. Compared to the total number of words in the corpus, oral patterns make up 7.7 per thousand words in Catalan, and 7.0 per thousand words in Spanish. Therefore, it seems that in our sample e-mails are to a large extent textual or written in nature, with little evidence of oral patterns. However, oral features deserve special attention due to their negative impact on translation quality.

In any case, the results indicate that successful application of MT to the online translation of e-mails in our environment would require customizing the MT system, taking into account the following: For both languages, unintentional deviations and terminology as well as intentional deviations are issues sufficiently important to require customizing the system. Regarding terminology, both domain terminology and user community terminology have to be dealt with, but the impact on translation is greater in the case of user community terminology, as most of the domain terminology is in English and does not usually need to be translated to be understood. The impact on translation of Internet vocabulary, which can be considered a special kind of terminology, is not significant, because the terms are widespread and commonly understood (note that they are usually English terms as well). Apart from terminology, the main sources of problems for translating e-mails in both languages are orthography and the lexicon; errors and deviations in syntax and pragmatics are scarcely significant. In terms of orthography, the most common problem is accentuation: 37.6%, in Catalan, and 46.8%, in Spanish, of all deviations involve accentuation, regardless of the fact that in Spanish this seems, for the most part, to be intentional (systematic lack of accentuation) and in Catalan, to be unintentional mistakes.

Conclusions and Future Directions

In this paper, we have presented an in-depth linguistic evaluation of a corpus of approximately 260 e-mails and 25,000 words, written in Catalan and Spanish, which came from four informal computer science newsgroups at the Universitat Oberta de Catalunya Virtual Campus. The messages were produced within a situation of bilingualism and language contact, where Spanish is progressively substituting for Catalan as the language of daily use.

The main goal of the study was to identify the linguistic characteristics of the e-mail register for our universe of study in order to assess their impact on machine translation, and based on our findings, to make recommendations on how to improve the quality of translation. The e-mails in our corpus were translated by an MT system, and the translations were evaluated, in order to formulate an improvement strategy: We classified and analyzed the features of the original messages that caused problems. The study is part of the INTERLINGUA project, which aims to adapt a system for online unsupervised translation of e-mails from Catalan to Spanish and vice versa, to avoid the marginalization of Catalan as a language for communication on the Internet. The main conclusions to be drawn from the study are the following: An implication of these findings is that the MT system must be fine-tuned in order to build software modules for the automatic correction of input, and of accents in particular. To the extent that the feasibility of incorporating minority languages such as Catalan into the multilingual Internet will depend on natural language processing, which usually deals with standard texts, difficulties might be expected for these languages in the future. At this point, an important question arises: Should MT systems bear the responsibility for correcting input in order to preserve users' spontaneity when writing, or should users be more careful in their use of language? While in theory the effort to write accurately might help increase the status of minority languages in CMC, the Internet is a medium that encourages non-standard writing and expressive innovation. In the environment studied, we cannot expect users always to write carefully, making the effort to look up words in the dictionary, use spelling and grammar checkers, etc. We think it unrealistic to rely on users to facilitate MT, and therefore believe that fine-tuning MT systems is essential for the survival of Catalan in newsgroups, chat and elsewhere on the Internet.

In the future, our efforts will be extended to purely monolingual environments, in particular to Spanish in non-bilingual territories, for example in Madrid. This will allow us to assess more precisely the effects of language interference when bilingual users write e-mails, i.e., by comparing the Spanish written by Catalans to the Spanish written by residents of Madrid. In this connection, the incidence of errors should be lower in monolingual environments: Unintentional non-standard features should play a less important role in the e-mail register for communities of users that do not have to deal with a situation of language contact. At the same time, communities that are not made up of university students or professionals (whether bilingual or monolingual) are expected to have a lower level of linguistic knowledge, which should lead to a higher incidence of errors. In addition, it might be hypothesized that e-mails written by other communities of Internet users, e.g., teenagers, would show greater use of intentional new forms of expressivity. Thus, the impact of expressivity on the characterization of the register would increase. Moreover, as mentioned above, teenagers would also be expected to make more errors due to a lack of awareness of the language, in a way that would confirm or even reinforce the tendencies shown in the present study.

The future of Catalan depends on its users but, on the multilingual internet, it also depends on the help of technological tools such as MT systems. We have seen that gaps in knowledge of the language are a very serious handicap in the usability of systems that aim to translate Catalan into other languages. We have also seen that these gaps are significantly present in one of the principal groups of Internet users, university students, and predict that they will also be significant in the other principal group, teenagers. It is further quite obvious that email is a medium that encourages linguistic innovations and creativity, and that this trend is not likely to change in the future. Consequently, we would encourage MT developers to fine-tune their systems to make them error-proof, yet also flexible enough to deal with the non-standard features typical of the e-mail register. Otherwise, if MT continues to focus only on controlled, standard, well written texts that contain no deviations from linguistic norms, Catalan is likely to disappear from the Internet in a few years.

Footnotes

1. Catalonia is an autonomous region in Spain that has self-governing powers guaranteed by the Spanish Constitution and the Catalan Statute of Autonomy. Although its actual political status is highly controversial, it can be seen as a kind of a "stateless nation" or "a nation within a nation". Moreover, some people use the word "Catalonia" to refer to the three Spanish autonomous regions (Catalonia, Balearic Islands, Valencia) and other territories in the South of France whose native language is Catalan. For simplicity's sake, in this paper, we will use the term "Catalonia" to refer to the strict autonomous region as defined and bounded by the Spanish Constitution (i.e., the region in northeast Spain whose capital is Barcelona).

2. The project is expected to be extended in the future to Catalan/English and Spanish/English.

3. For further information on the sociolinguistic situation in Catalonia, see Pujolar (2000) and Strubell and Hall (1992).

4. We term it UOC-Catalonia, as the institution has recently opened a line of studies in Spanish for the rest of Spain and Latin America.

5. See Danet (2001, ch.1) for a more comprehensive and in-depth review of recent literature on this topic.

6. It is debatable whether these features are exclusive to e-mail or whether they also characterize other types of text-based CMC.

7. This corpus is a sub-set of the one used for the analysis of code-switching in the first phase of this study.

Acknowledgments

We are grateful to Susan Herring, Brenda Danet and the anonymous reviewers of this paper for their useful critical comments and observations. Needless to say, all remaining errors are our own.

The research presented in this paper has been partially funded by the project IR-226 (Internet Interdisciplinary Institute, IN3/UOC: project INTERLINGUA).

References

ÀLATAC. (2003). Website of the Servei de Llengües i Terminologia de la Universitat Politècnica de Catalunya i l'Associació del Voluntariat Lingüístic. Retrieved July 18, 2003 from http://www.upc.es/slt/alatac/cat/dades/catala.html.

Alonso, A., Folguerà, R., & Tebé, C. (2000). Del tecnolecte al sociolecte: consideracions sobre l'argot tècnic en català. In M. Torres, Ll. Jardí, N. Alturo, Ll. Payrató & F.X. Vila (Eds.) CMO-Cat I Jornada sobre comunicació mediatitzada per ordinador en català. Retrieved July 18, 2003 from http://www.ub.es/lincat/cmo-cat/tebe-alonso-folguera.htm.

Badia, J., Bertran, C., & Castells, A. (2001). És possible viure en català? Barcelona: Angle.

Biber, D. (1988). Variation across speech and writing. Cambridge: Cambridge University Press.

Castells, M., & Díaz de la Isla, M. I. (2001). Diffusion and uses of Internet in Catalonia and Spain. Project Internet Catalonia (PIC) Working Paper Series, PICWP 1201. Barcelona, Spain: Universitat Oberta de Catalunya / IN3. Retrieved July 18, 2003 from http://www.uoc.edu/in3/wp/picwp1201/.

Cerdà, R. (2001). Castellano y catalán en Cataluña y las Islas Baleares. II Congreso Internacional de la Lengua Española. 2001 Valladolid, Spain: Proceedings. Retrieved July 18, 2003 from http://cvc.cervantes.es/obref/congresos/valladolid.

CIS: Centro de Investigaciones Sociológicas. (1998). Uso de lenguas en comunidades bilingües (II): Cataluña. In Catálogo del banco de datos del Centro de Investigaciones Sociológicas, estudio 2298, Madrid, Spain. Retrieved July 18, 2003 from http://www.cis.es/estudio.asp?nest=2298.

Climent, S., Moré, J., & Oliver, A. (2003). Building an environment for unsupervised automatic email translation. EAMT-CLAW 2003, Joint Conference combining the 7th international workshop of the European Association for Machine Translation and the 4th Controlled Language Applications Workshop, Dublin: Proceedings. Retrieved July 18, 2003 from http://www.uoc.edu/in3/dt/20292/index.html.

Danet, B. (2001). Cyberpl@y: Communicating online. Oxford: Berg.

EC: European Commission. (1998). The euromap report. Luxembourg: Linglink.

Fais, L., & Ogura, K. (2001). Discourse issues in the translation of Japanese e-mail. Conference of the Pacific Association for Computational Linguistics, PACLING 2001, Kitakyushu, Japan: Proceedings. Retrieved March 27, 2003 from http://afnlp.org/pacling2001/pdf/fais.pdf.

Ferrara, K., Brunner, H., & Whitemore, G. (1990). Interactive written discourse as an emergent register. Written Communication, 8, 8-34.

Green, R. (1977). Analysis of errors. Commission of the European Communities (CEC) memorandum, October, 5+5 p. Luxembourg.

Herring, S. C. (1999). Interactional coherence in CMC. In T. Erickson (Ed.), Journal of Computer-Mediated Communication, 4 (4), special issue on Persistent Conversation. Retrieved March 27, 2003 from http://jcmc.indiana.edu/vol4/issue4/herring.html.

Herring, S. C. (2001). Computer-mediated discourse. In D. Tannen, D. Schiffin & H. Hamilton (Eds.), Handbook of discourse analysis (pp. 612-634). Oxford: Blackwell.

IDESCAT, Institut d'Estadística de Catalunya. (2001). Interactive Statistics Request performed at the IDESCAT web site, http://www.idescat.es. Retrieved March 27, 2003 from http://www.idescat.es/scripts/sqldequavi.dll?TC=444&V0=8&V1=6.

ISLE: International Standards for Language Engineering. (2000). The ISLE classification of machine translation evaluations. Retrieved March 27, 2003 from http://www.isi.edu/natural-language/mteval.

Mehsching, G. (2000). The internet as a rescue tool of endangered languages: Sardinian. Retrieved July 29, 2003 from http://www.gaia.es/multilinguae/pdf/Guido.PDF.

Murray, D. E. (2000). Protean communication: The language of computer-mediated communication. TESOL Quarterly, 34 (3), 397-421.

Payà, M. (2000). Com responem els missatges de correu electrònic? Noves formes de diàleg. In M. Torres, Ll. Jardí, N. Alturo, Ll. Payrató & F.X. Vila (Eds.), CMO-Cat I jornada sobre comunicació mediatitzada per ordinador en català. Retrieved July 18, 2003 from http://www.ub.es/lincat/cmo-cat/paya.htm.

Pujolar, J. (2000). Gender, heteroglossia and power. A sociolinguistic study of youth culture. Berlin: Mouton de Gruyter.

Siguán, M. (1999). Conocimiento y uso de las lenguas. Madrid, Spain: Centro de Investigacions Sociológicas, CIS, 22.

Siguán, M. (2001). El Español como lengua en contacto con otras lenguas. II Congreso Internacional de la Lengua Española. Valladolid, Spain: Proceedings. Retrieved March 27, 2003 from http://cvc.cervantes.es/obref/congresos/valladolid.

Strubell, M., & Hall, J. (1992). Problems and prospects of small linguistic societies - Catalonia. EMI Education Media International, (29) 1, 26-37.

Turell, Mª T. (Ed.) (2001). Multilingualism in Spain: Sociolinguistic and psycholinguistic aspects of linguistic minority groups. Clevedon: Multilingual Matters.

Van Slype, G. (1979). Critical study of methods for evaluating the quality of machine translation. Commission of the European Communities Directorate General Scientific and Technical Information and Information Management Report BR 19142. Retrieved March 27, 2003 from http://www.ling.ed.ac.uk/~beatrice/bibliography.htm.

Warschauer, M. (2000). Language, identity, and the Internet. In B. Kolko, L. Nakamura, & G. Rodman (Eds.), Race in cyberspace. New York: Routledge. Also accessed July 29, 2003 from http://www.gse.uci.edu/markw/lang.htm. .

Yates, J. A., & Orlikowski, W. J. (1993). Knee-jerk anti-LOOPism and other e-mail phenomena: Oral, written, and electronic patterns in computer-mediated communication. MIT Sloan School Working Paper 3578-93, Center for Coordination Science Technical Report 150. Retrieved March 27, 2003 from http://ccs.mit.edu/papers/CCSWP150.html.

About the Authors

Salvador Climent has a master's degree in Linguistics from the Universitat de Barcelona, a postgraduate master in Computational Linguistics, and a Ph.D. in Language and Cognitive Science from the same University. He is a professor of Humanities and Catalan Philology at the Universitat Oberta de Catalunya.
Address: Universitat Oberta de Catalunya. Estudis d'Humanitats i Filologia catalana. Av. Tibidabo 39-43. 08035 Barcelona.

Joaquim Moré has a master's degree in English Philology from the Universitat de Barcelona and a postgraduate master's degree in Computational Linguistics from the same University. He is a researcher at the Internet Interdisciplinary Institute.
Address: Universitat Oberta de Catalunya. IN3. Av. Tibidabo 39-43. 08035 Barcelona.

Antoni Oliver has a degree in Telecommunications Engineering from the Universitat Politècnica de Catalunya and a master's degree in Slavic Philology from the Universitat de Barcelona. He is a researcher at the Internet Interdisciplinary Institute.
Address: Universitat Oberta de Catalunya. IN3. Av. Tibidabo 39-43. 08035 Barcelona.

Míriam Salvatierra has a master's degree in Catalan Philology from the Universitat de Barcelona. She is a Language Technician at the Linguistic Service at the Universitat Oberta de Catalunya.
Address: Universitat Oberta de Catalunya. Servei Lingüístic. Diputació 219. 08011 Barcelona.

Imma Sànchez has a master's degree in Catalan Philology from the Universitat de Barcelona and a postgraduate degree in Applied and Computational Linguistics from the same University. She is the director of the Linguistic Service at the Universitat Oberta de Catalunya.
Address: Universitat Oberta de Catalunya. Servei Lingüístic. Diputació 219. 08011 Barcelona.

Mariona Taulé has a master's degree and a Ph.D. in Linguistics from the Universitat de Barcelona. She is professor of Linguistics at the same University and at the Universitat Oberta de Catalunya.
Address: Universitat de Barcelona. Departament de Lingüística. Gran Via de les Corts Catalanes 585. 08007 Barcelona.

Lluïsa Vallmanya has a degree in Catalan Language Teaching and a master's degree in Catalan Philology. She is a Language Technician at the Linguistic Service at the Universitat Oberta de Catalunya.
Address: Universitat Oberta de Catalunya. Servei Lingüístic. Diputació 219. 08011 Barcelona.


©Copyright 2003 Journal of Computer-Mediated Communication