Text to text applications include machine translation, automated summarization, question answering, and other similar applications where a machine carries out the function of understanding some kind of input information, and generating text. The input information is often “text”, but more generally, can be any kind of information that is received and understandable by the machine.
Conventional text to text applications use heterogeneous methods for implementing the generation phase. Machine translation often produces sentences using application-specific decoders that are based on work that was conducted on speech recognition. Automated summarization produces abstracts using task specific strategies.
Machine translation systems rely on training that is carried out based on corresponding, or “parallel” information that exists in both of two languages. The information in the two languages can be from many sources. Sometimes, it is known that the contents of two documents represent the same information.
The internet is a source of information. Documents on the Internet are often available in multiple different languages. However, it may be difficult to identify mutual translations within the many different web pages on the Internet. Comparing all documents within the document pool using conventional systems would require a number of computations that scales with the square of the number of document pairs.
For example, each English language page can be compared with every known French language page, to determine the best match. This naive system would take extreme computation times to identify the training pairs.
Philip Resnik has suggested a method which identifies parallel documents by producing pairs of similar URLs which are presumed to be in different languages. For example, if one URL says “En”, and another URL is similar but differs only by stating “FR”, then these are presumed to be parallel URLs.
Not all Web documents are in this form, and Resnik's system is quite specific to web pages which have that specific kinds of URLs.
The present application teaches a system that forms a similarity measure that returns a score given a document pair. Techniques are disclosed which scale n*log n with the number of documents.
One aspect forms a reduced-size version of the document that is associated with the document contents, and compares that reduced size version, with comparably reduced sized versions in other languages. The reduced size document can be a document fingerprint.
Another aspect compares the documents using a probabilistic shuffling technique, where the documents and contents are mixed, and then compared to some, but not all, information about other documents. The shuffling may be carried out numerous times, in order to obtain a best match.
The drawings show:
The general structure and techniques, and more specific embodiments which can be used to effect different ways of carrying out the more general goals are described herein.
A processor is assumed to have access to various sources 105. The sources may be parallel corpora of multiple language information. Specifically, the sources may include translation memories, probabilistic and non-probabilistic word- and phrase-based dictionaries, glossaries, Internet information, parallel corpora in multiple languages, non-parallel corpora in multiple languages having similar subject matter, and human-created translations. The processor creates training data 110.
Speech engine 120 carries out a text-to-text application based on the training data.
The present application teaches a system of identifying mutual translations within a collection of documents such as 105. The documents are assumed to be in first and second languages.
A first embodiment describes the first and second languages as being English and French. It should be understood, however, that any first and second languages could be used. The language information is used to train a machine based text to text system. That system can be machine translation, automated summarization, speech recognition, or any other machine application.
Data from the Web can be gathered by focused crawling. Alternatively, other data can be obtained. The data includes a collection of information in first and second languages that does not necessarily have any subject matter connection. This data is used as the input to the system.
The processing computer operates to find mutual translations according to the flowchart of
At 210, reduced size versions of the documents are created for both the native and translated documents. The reduced size version has parts that are associated with the document contents. The reduced size document can be a document fingerprint. The fingerprint has “keys” that relate to words and their placement in the dictionaries. In effect, this summarizes concisely information about the words contained in the document.
Different techniques of forming fingerprints may be used, and one specific technique is described herein with reference to
The keys collectively form a fingerprint. A typical system of this type may use 128 different dictionaries, and hence the fingerprint, shown in 315 is formed of 128 different keys. Each document will form a unique set of keys, and conversely, the keys effectively form a signature that allows identification of the document. Any other type signature which identifies the document can alternatively be used. At 220, each of the native and translated documents is compared to its neighboring document, that is not to all documents in the database, but only to specified neighboring documents. The comparison may use, for example, a fast Hamming matching. The comparison may only be to the left and right neighbors, or may alternatively be to 2-5 left and right nearest neighbors, or to some other number of neighbors. The Hamming distance is found at 225 and represents how many pieces of the pair of fingerprints do not match.
Even a document and its identical translation would not match exactly because of imperfections in the translator, for example. The Hamming distance indicates the amount by which the fingerprints do not match.
At 230, a shuffle is carried out, in which the order of the keys within the native and translated fingerprints are shuffled randomly. After shuffling, the documents are sorted at 235, according to fingerprints. The documents are again compared to their nearest neighbor(s) at 225. Flow continues until a desired match is obtained. The output is the closest neighbor at 240.
The shuffle operation uses statistical properties to find the nearest neighbor. For a database with 1,000 documents, for example, the shuffle can find the nearest neighbor after approximately 50 shuffles.
The SHUFFLE process is done so that the keys can be sorted in a way that brings similar items “nearby”, so that they get compared.
For example, consider the following two 5-key signatures:
These two docs are quite similar because they differ only in one key (ie, 0 vs 9).
However, the ordering of the documents may be very different, depending on the key order. A worst-case shuffle, for example, may lead to the following key re-ordering:
When documents are sorted according to their keys and according to this worst case scenario, doc-1 & doc-2 are likely to be very far apart. An example sorting might be:
In contrast, a best-case shuffle will put the like keys in agreement, for example, a best case shuffle might be:
In this case, after sorting, the documents will be very close.
Another embodiment is described with reference to the flowchart of
At 400, each document in the collection is analyzed according to an assessment vector technique. The analysis may look for any category or feature within each document. For example, the assess operation at 400 may maintain a detection of the number of times that a specified word is used, and keep counts of those multiple uses. The analyzed information is used to form vectors indicative of the documents. In this embodiment, the vectors form the reduced versions.
The vectors can be native, or may use a translation technique. For example, a word frequency vector can be used for English documents, while a modified word frequency vector can be used place the words from the French document into the English space.
At 420, the vectors are compared and shuffled at 430 using similar techniques to those in a previous embodiment.
According to exemplary embodiments, a processor may determine reduced size versions of documents such as those that may be included in a database. A processor may also compare the reduced size versions to determine documents that represent similar information. Additionally, a text-to-text application module may use documents that represent similar information to train a text-to-text application. The text-to-text application may be a machine translation system in some embodiments. Furthermore, the text-to-text application may carry out a rough translation of documents to form a group of translated documents.
Although only a few embodiments have been disclosed in detail above, other embodiments are possible and are intended to be encompassed within this specification. The specification describes specific examples to accomplish a more general goal that may be accomplished in other way. This disclosure is intended to be exemplary, and the claims are intended to cover any modification or alternative which might be predictable to a person having ordinary skill in the art. For example, the above techniques can be used with other sources of information, other languages, and other signature techniques.
Also, only those claims which use the words “means for” are intended to be interpreted under 35 USC 112, sixth paragraph. Moreover, no limitations from the specification are intended to be read into any claims, unless those limitations are expressly included in the claims.
This invention was made with government support under Contract No. N66001-00-1-8914 awarded by the Space and Naval Warfare Systems Command. The U.S. government has certain rights in the claimed inventions.
Number | Name | Date | Kind |
---|---|---|---|
4502128 | Okajima et al. | Feb 1985 | A |
4599691 | Sakaki et al. | Jul 1986 | A |
4787038 | Doi et al. | Nov 1988 | A |
4814987 | Miyao et al. | Mar 1989 | A |
4942526 | Okajima et al. | Jul 1990 | A |
5146405 | Church | Sep 1992 | A |
5181163 | Nakajima et al. | Jan 1993 | A |
5212730 | Wheatley et al. | May 1993 | A |
5267156 | Nomiyama | Nov 1993 | A |
5311429 | Tominaga | May 1994 | A |
5432948 | Davis et al. | Jul 1995 | A |
5477451 | Brown et al. | Dec 1995 | A |
5510981 | Berger et al. | Apr 1996 | A |
5644774 | Fukumochi et al. | Jul 1997 | A |
5696980 | Brew | Dec 1997 | A |
5724593 | Hargrave III et al. | Mar 1998 | A |
5761631 | Nasukawa | Jun 1998 | A |
5781884 | Pereira et al. | Jul 1998 | A |
5794178 | Caid et al. | Aug 1998 | A |
5805832 | Brown et al. | Sep 1998 | A |
5848385 | Poznanski et al. | Dec 1998 | A |
5867811 | O'Donoghue | Feb 1999 | A |
5870706 | Alshawi | Feb 1999 | A |
5903858 | Saraki | May 1999 | A |
5987404 | Della Pietra et al. | Nov 1999 | A |
5991710 | Papineni et al. | Nov 1999 | A |
6031984 | Walser | Feb 2000 | A |
6032111 | Mohri | Feb 2000 | A |
6092034 | McCarley et al. | Jul 2000 | A |
6119077 | Shinozaki | Sep 2000 | A |
6131082 | Hargrave III et al. | Oct 2000 | A |
6182014 | Kenyon et al. | Jan 2001 | B1 |
6205456 | Nakao | Mar 2001 | B1 |
6223150 | Duan et al. | Apr 2001 | B1 |
6236958 | Lange et al. | May 2001 | B1 |
6278967 | Akers et al. | Aug 2001 | B1 |
6285978 | Bernth et al. | Sep 2001 | B1 |
6289302 | Kuo | Sep 2001 | B1 |
6304841 | Berger et al. | Oct 2001 | B1 |
6311152 | Bai et al. | Oct 2001 | B1 |
6317708 | Witbrock et al. | Nov 2001 | B1 |
6360196 | Poznanski et al. | Mar 2002 | B1 |
6389387 | Poznanski et al. | May 2002 | B1 |
6393388 | Franz et al. | May 2002 | B1 |
6393389 | Chanod et al. | May 2002 | B1 |
6415250 | van den Akker | Jul 2002 | B1 |
6460015 | Hetherington et al. | Oct 2002 | B1 |
6502064 | Miyahira et al. | Dec 2002 | B1 |
6782356 | Lopke | Aug 2004 | B1 |
6810374 | Kang | Oct 2004 | B2 |
6904402 | Wang et al. | Jun 2005 | B1 |
7107215 | Ghali | Sep 2006 | B2 |
7113903 | Riccardi et al. | Sep 2006 | B1 |
7197451 | Carter et al. | Mar 2007 | B1 |
7356457 | Pinkham et al. | Apr 2008 | B2 |
20020078091 | Vu et al. | Jun 2002 | A1 |
20020188438 | Knight et al. | Dec 2002 | A1 |
20020198701 | Moore | Dec 2002 | A1 |
20040030551 | Marcu et al. | Feb 2004 | A1 |
20050228643 | Munteanu et al. | Oct 2005 | A1 |
20060015320 | Och | Jan 2006 | A1 |
20060142995 | Knight et al. | Jun 2006 | A1 |
20060150069 | Chang | Jul 2006 | A1 |
20090083023 | Foster et al. | Mar 2009 | A1 |
Number | Date | Country |
---|---|---|
0469884 | Feb 1992 | EP |
0715265 | Jun 1996 | EP |
0933712 | Aug 1999 | EP |
07244666 | Jan 1995 | JP |
10011447 | Jan 1998 | JP |
11272672 | Oct 1999 | JP |
Number | Date | Country | |
---|---|---|---|
20070033001 A1 | Feb 2007 | US |