Baby-Steps Towards Building a Spanglish Language Model

Spanglish is the simultaneous use, or alternating of both, traditional Spanish and English within the same conversational event. This interlanguage is commonly used in U.S. populations with large percentages of Spanish speakers. Despite the popularity of this dialect, and the wide spread of automated voice systems, currently there are no spoken dialog applications that can process Spanglish. In this paper we present the first attempt towards creating a Spanglish language model. 1 What is Spanglish? Spanglish has existed for a long time, but has not been formally recognized as a language, nor has it been classified as a particular linguistic phenomenon. This interlanguage is more of a continuum of the mix between English and Spanish. From a linguistic point of view, it is difficult to decide what to consider Spanglish. It is debatable whether to consider Spanglish as an interlanguage, a pidgin, or a creole language. An interlanguage is a language that is often spoken between linguistic borders [1]; Spanglish does not fit this category, as it is also spoken in areas where no such borders exist, New York City being an example of this. A pidgin is a communication system created when people communicate despite their lack of knowledge in the other language [1]; this might explain its origin, but it certainly does not apply to its use, as most of the Spanglish speakers are bilingual. A creole language originates when a community adopts a pidgin as their primary source for communication [1]; a fragment of Spanglish speakers fall under this category since they cannot use traditional English or Spanish because of lack of proper training, but this cannot be generalized to all the Spanglish speakers, a large percentage of Spanglish speakers are bilingual who can express themselves in either of the traditional languages. The origins of Spanglish in the U.S. are attributed, to a large extent, to socio-historical circumstances. The Mexican-American war, which according to history, started with the annexation of Texas to the U.S., resulted in Mexico ceding the territories of California and New Mexico to the U.S. in the mid eighteen hundreds. For many years Spanish speakers were going back and forth across these regions maintaining contact with English speakers. Many years later, the U.S. experienced a considerable immigration from Spanish speaking countries like Mexico, Cuba, Venezuela, Colombia and even Spain. In recent years, the flow of immigrants from Spanish speaking countries has not ceased to occur. In addition to this, the constant contact among the border cities between the U.S. and Mexico certainly has had influence on the proliferation of Spanglish. In this paper we report results from building a Language Model (LM) with a small Spanglish corpus we collected. To the best of our knowledge, we are the first attempting to build a LM for Spanglish. Such LM is one of the first steps towards advancing the state-of-the-art regarding the automated processing of interlanguages, an achievement that will open the road for exploring interesting research avenues and applications. A good example is the possibility for building an automated speech recognizer for spoken dialog systems capable of processing requests from Spanglish speakers. We present here evaluation results of the language model, and although they show the language model to be weak, the results are promising. We will continue working on gathering more data to improve the corpus. However, the corpus already represents a valuable asset for deeper analysis of bilingualism. It will allow a statistical analysis that can support a formal characterization of Spanglish. The next section describes some of the most salient features of Spanglish. 2 Linguistic Features of Spanglish In the linguistic, sociolinguistic, psychology, and psycholinguistic literature, bilingualism and the inherent phenomena it exhibits has been studied for nearly a century [7, 8, 11–13, 16, 20]. Despite the numerous previous studies of linguistic characteristics of bilingualism, there is not a clear consensus on the use of concepts related to the language alternation patterns in bilingual speakers. The alternation of languages within a sentence is known as code-mixing, but it has also been refereed as intrasentential code-switching, and intrasentential alternation [1, 10, 18]. Alternation across sentence boundaries is known as intersentential code-switching, or just code-switching. Yet there is another alternation mode defined as borrowing, which consists on adopting words, or idiomatic expressions, of a foreign language, usually modifying the original word, or expression, to suit the grammar or morphology of the receiving language [19]. In this paper we present Ardila’s classification of Spanglish characteristics into two groups: shallow and deep phenomena. From his definition, shallow phenomena encompass code-mixing and code-switching; these are the linguistic features of Spanglish that can be easily spotted by humans. In contrast, deep phenomena includes, among other things, the transformation of Spanish to approximate English; the transformations can be so subtle that they are harder to detect, even for speakers of traditional Spanish, and include false cognates, also known as false friends. For our research purpose we are interested mostly in shallow phenomena of Spanglish, thus, the following subsections are focused on this type of features. The interested reader can find more information regarding the deep phenomena in [1].