section string | sentence string | label int64 |
|---|---|---|
ABSTRACT | A major obstacle to the construction of a probabilistic translation model is the lack of large parallel corpora. | 0 |
ABSTRACT | In this paper we first describe a parallel text mining system that finds parallel texts automatically on the Web. | 0 |
ABSTRACT | The generated Chinese-English parallel corpus is used to train a probabilistic translation model which translates queries for Chinese-English cross-language information retrieval (CLIR). | 0 |
ABSTRACT | We will discuss some problems in translation model training and show the preliminary CUR results. | 0 |
INTRO | Parallel texts have been used in a number of studies in computational linguistics. | 0 |
INTRO | defined a series of probabilistic translation models for MT purposes. | 1 |
INTRO | While people may question the effectiveness of using these models for a full-blown MT system, the models are certainly valuable for developing translation assistance tools. | 0 |
INTRO | For example, we can use such a translation model to help complete target text being drafted by a human translator. | 1 |
INTRO | Another utilization is in cross-language information retrieval (CUR) where queries have to be translated from one language to another language in which the documents are written. | 0 |
INTRO | In CUR, the quality requirement for translation is relatively low. | 0 |
INTRO | For example, the syntactic aspect is irrelevant. | 0 |
INTRO | Even if the translated word is not a true translation but is strongly related to the original query, it is still helpful. | 0 |
INTRO | Therefore, CUR is a suitable application for such a translation model. | 0 |
INTRO | However, a major obstacle to this approach is the lack of parallel corpora for model training. | 0 |
INTRO | Only a few such corpora exist, including the Hansard English-French corpus and the HKUST EnglishChinese corpus. | 1 |
INTRO | In this paper, we will describe a method which automatically searches for parallel texts on the Web. | 0 |
INTRO | We will discuss the text mining algorithm we adopted, some issues in translation model training using the generated parallel corpus, and finally the translation model's performance in CUR. | 0 |
method | The PTMiner system is an intelligent Web agent that is designed to search for large amounts of parallel text on the Web. | 0 |
method | The mining algorithm is largely language independent. | 0 |
method | It can thus be adapted to other language pairs with only minor modifications. | 0 |
method | We take advantage of the huge number of Web sites indexed by existing search engines in determining candidate sites. | 0 |
method | This is done by submitting some particular requests to the search engines. | 0 |
method | The requests are determined according to the following observations. | 0 |
method | In the sites where parallel text exists, there are normally some pages in one language containing links to the parallel version in the other language. | 0 |
method | These are usually indicated by those links' anchor texts 1. | 0 |
method | For example, on some English page there may be a link to its Chinese version with the anchor text "Chinese Version" or "in Chinese". | 0 |
method | The same phenomenon can be observed on Chinese pages. | 0 |
method | Chances are that a site with parallel texts will contain such links in some of its documents. | 0 |
method | This fact is used as the criterion in searching for candidate sites. | 0 |
method | Therefore, to determine possible sites for EnglishChinese parallel texts, we can request an English document containing the following anchor: anchor: "english version" ["in english", ...]. | 0 |
method | Similar requests are sent for Chinese documents. | 0 |
method | From the two sets of pages obtained by the above queries we extract two sets of Web sites. | 0 |
method | The union of these two sets constitutes then the candidate sites. | 0 |
method | That is to say, a site is a candidate site when it is found to have either an English page linking to its Chinese version or a Chinese page linking to its English version. | 0 |
method | We now assume that a pair of parallel texts exists on the same site. | 0 |
method | To search for parallel pairs on a site, PTMiner first has to obtain all (or at least part of) the HTML file names on the site. | 0 |
method | From these names pairs are scanned. | 0 |
method | It is possible to use a Web crawler to explore the candidate sites completely. | 0 |
method | However, we can take advantage of the search engines again to accelerate the process. | 0 |
method | As the first step, we submit the following query to the search engines: host: hostname to fetch the Web pages that they indexed from this site. | 0 |
method | If we only require a small amount of parallel texts, this result may be sufficient. | 0 |
method | For our purpose, however, we need to explore the sites more thoroughly using a host crawler. | 0 |
method | Therefore, we continue our search for files with a host crawler which uses the documents found by the search engines as the starting point. | 0 |
method | A host crawler is slightly different from a Web crawler. | 0 |
method | Web crawlers go through innumerable pages and hosts on the Web. | 0 |
method | A host crawler is a Web crawler that crawls through documents on a given host only. | 0 |
method | A breadth-first crawling algorithm is applied in PTMiner as host crawler. | 0 |
method | The principle is that when a link to an unexplored document on the same site is found in a document, it is added to a list that will be explored later. | 0 |
method | In this way, most file names from the candidate sites are obtained. | 0 |
method | After collecting file names for each candidate site, the next task is to determine the parallel pairs. | 0 |
method | Again, we try to use some heuristic rules to guess which files may be parallel texts before downloading them. | 0 |
method | The rules are based on external features of the documents. | 0 |
method | By external feature, we mean those features which may be known without analyzing the contents of the file, such as its URL, size, and date. | 0 |
method | This is in contrast with the internal features, such as language, character set, and HTML structure, which cannot be known until we have downloaded the page and analyzed its contents. | 0 |
method | The heuristic criterion comes from the following observation: We observe that parallel text pairs usually have similar name patterns. | 0 |
method | The difference between the names of two parallel pages usually lies in a segment which indicates the language. | 0 |
method | For example, "file-ch.html" (in Chinese) vs. "file-en.html" (in English). | 0 |
method | The difference may also appear in the path, such as ".../chinese/.../file.html" vs. | 0 |
method | glish/.../file.html" . The name patterns described above are commonly used by webmasters to help organize their sites. | 0 |
method | Hence, we can suppose that a pair of pages with this kind of pattern are probably parallel texts. | 0 |
method | First, we establish four lists for English prefixes, English suffixes, Chinese prefixes and Chinese suffixes. | 0 |
method | For example: English Prefix = {e, en, e_, en_, e—, en—, ...}. | 0 |
method | For each file in one language, if a segment in its name corresponds to one of the language affixes, several new names are generated by changing the segment to the possible corresponding affixes of the other language. | 0 |
method | If a generated name corresponds to an existing file, then the file is considered as a candidate parallel document of the original file. | 0 |
method | Next, we further examine the contents of the paired files to determine if they are really parallel according to various external and internal features. | 0 |
method | This may further improve the pairing precision. | 0 |
method | The following methods have been implemented in our system. | 0 |
method | Parallel files often have similar file lengths. | 0 |
method | One simple way to filter out incorrect pairs is to compare the lengths of the two files. | 0 |
method | The only problem is to set a reasonable threshold that will not discard too many good pairs, i.e. balance recall and precision. | 0 |
method | The usual difference ratio depends on the language pairs we are dealing with. | 0 |
method | For example, ChineseEnglish parallel texts usually have a larger difference ratio than English-French parallel texts. | 0 |
method | The filtering threshold had to be determined empirically, from the actual observations. | 0 |
method | For Chinese-English, a difference up to 50% is tolerated. | 0 |
method | It is also obvious that the two files of a pair have to be in the two languages of interest. | 0 |
method | By automatically identifying language and character set, we can filter out the pairs that do not satisfy this basic criterion. | 0 |
method | Some Web pages explicitly indicate the language and the character set. | 0 |
method | More often such information is omitted by authors. | 0 |
method | We need some language identification tool for this task. | 0 |
method | SILC is a language and encoding identification system developed by the RALI laboratory at the University of Montreal. | 0 |
method | It employs a probabilistic model estimated on tri-grams. | 0 |
method | Using these models, the system is able to determine the most probable language and encoding of a text. | 1 |
method | In the STRAND system, the candidate pairs are evaluated by aligning them according to their HTML structures and computing confidence values. | 1 |
method | Pairs are assumed to be wrong if they have too many mismatching markups or low confidence values. | 0 |
method | Comparing HTML structures seems to be a sound way to evaluate candidate pairs since parallel pairs usually have similar HTML structures. | 0 |
method | However, we also noticed that parallel texts may have quite different HTML structures. | 0 |
method | One of the reasons is that the two files may be created using two HTML editors. | 0 |
method | For example, one may be used for English and another for Chinese, depending on the language handling capability of the editors. | 0 |
method | Therefore, caution is required when measuring structure difference numerically. | 0 |
method | Parallel text alignment is still an experimental area. | 0 |
method | Measuring the confidence values of an alignment is even more complicated. | 0 |
method | For example, the alignment algorithm we used in the training of the statistical translation model produces acceptable alignment results but it does not provide a confidence value that we can "confidently" use as an evaluation criterion. | 0 |
method | So, for the moment this criterion is not used in candidate pair evaluation. | 0 |
method | In this section, we describe the results of our parallel text mining and translation model training. | 0 |
method | Using the above approach for Chinese-English, 185 candidate sites were searched from the domain hk. | 0 |
method | We limited the mining domain to hk because Hong Kong is a bilingual English-Chinese city where high quality parallel Web sites exist. | 0 |
method | Because of the small number of candidate sites, the host crawler was used to thoroughly explore each site. | 0 |
method | The resulting corpus contains 14820 pairs of texts including 117.2Mb Chinese texts and 136.5Mb English texts. | 0 |
method | The entire mining process lasted about a week. | 0 |
method | Using length comparison and language identification, we refined the precision of the corpus to about 90%. | 0 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.