nakov paskaleva cognate or false friend ask the we

Cognate or False Friend? Ask the Web! : Cognate or False Friend? Ask the Web! Svetlin Nakov, Sofia University "St. Kliment Ohridski" Preslav Nakov, University of California, Berkeley Elena Paskaleva, Bulgarian Academy of Sciences A Workshop on Acquisition and Management of Multilingual Lexicons


Introduction : Introduction Cognates and false friends Cognates are pair of words in different languages that sound similar and are translations of each other False friends are pairs of words in two languages that sound similar but differ in their meanings The problem Design an algorithm that can distinguish between cognates and false friends


Cognates and False Friends : Cognates and False Friends Examples of cognates ден in Bulgarian = день in Russian (day) idea in English = идея in Bulgarian (idea) Examples of false friends майка in Bulgarian (mother) ≠ майка in Russian (vest) prost in German (cheers) ≠ прост in Bulgarian (stupid) gift in German (poison) ≠ gift in English (present)


The Paper in One Slide : The Paper in One Slide Measuring semantic similarity Analyze the words local contexts Use the Web as a corpus Similarities contexts  similar words Context translation  cross-lingual similarity Evaluation 200 pairs of words 100 cognates and 100 false friends 11pt average precision: 95.84%


Contextual Web Similarity : Contextual Web Similarity What is local context? Few words before and after the target word The words in the local context of given word are semantically related to it Need to exclude the stop words: prepositions, pronouns, conjunctions, etc. Stop words appear in all contexts Need of sufficiently big corpus Same day delivery of fresh flowers, roses, and unique gift baskets from our online boutique. Flower delivery online by local florists for birthday flowers.


Contextual Web Similarity : Contextual Web Similarity Web as a corpus The Web can be used as a corpus to extract the local context for given word The Web is the largest possible corpus Contains big corpora in any language Searching some word in Google can return up to 1 000 excerpts of texts The target word is given along with its local context: few words before and after it Target language can be specified


Contextual Web Similarity : Contextual Web Similarity Web as a corpus Example: Google query for "flower"


Contextual Web Similarity : Contextual Web Similarity Measuring semantic similarity For given two words their local contexts are extracted from the Web A set of words and their frequencies Semantic similarity is measured as similarity between these local contexts Local contexts are represented as frequency vectors for given set of words Cosine between the frequency vectors in the Euclidean space is calculated


Contextual Web Similarity : Contextual Web Similarity Example of context words frequencies word: flower word: computer


Contextual Web Similarity : Contextual Web Similarity Example of frequency vectors Similarity = cosine(v1, v2) v1: flower v2: computer


Cross-Lingual Similarity : Cross-Lingual Similarity We are given two words in different languages L1 and L2 We have a bilingual glossary G of translation pairs {p ∈ L1, q ∈ L2} Measuring cross-lingual similarity: We extract the local contexts of the target words from the Web: C1 ∈ L1 and C2 ∈ L2 We translate the context We measure distance between C1* and C2


Reverse Context Lookup : Reverse Context Lookup Local context extracted from the Web can contain arbitrary parasite words like "online", "home", "search", "click", etc. Internet terms appear in any Web page Such words are not likely to be associated with the target word Example (for the word flowers) "send flowers online", "flowers here", "order flowers here" Will the word "flowers" appear in the local context of "send", "online" and "here"?


Reverse Context Lookup : Reverse Context Lookup If two words are semantically related both should appear in the local contexts of each other Let #{x,y} = number of occurrences of x in the local context of y For any word w and a word from its local context wc, we define their strength of semantic association p(w,wc) as follows: p(w, wc) = min{ #(w, wc), #(wc,w) } We use p(w,wc) as vector coordinates when measuring semantic similarity


Web Similarity Using Seed Words : Web Similarity Using Seed Words Adaptation of the Fung&Yee'98 algorithm* We have a bilingual glossary G: L1  L2 of translation pairs and target words w1, w2 We search in Google the co-occurrences of the target words with the glossary entries Compare the co-occurrence vectors for each {p,q} ∈ G compare max (google#("w1 p") and google#("p w1")) with max (google#"w2 q") and google#("q w2")) * P. Fung and L. Y. Yee. An IR approach for translating from nonparallel, comparable texts. In Proceedings of ACL, volume 1, pages 414–420, 1998


Evaluation Data Set : Evaluation Data Set We use 200 Bulgarian/Russian pairs of words: 100 cognates and 100 false friends Manually assembled by a linguist Manually checked in several large monolingual and bilingual dictionaries Limited to nouns only


Experiments : Experiments We tested few modifications of our contextual Web similarity algorithm Use of TF.IDF weighting Preserve the stop words Use of lemmatization of the context words Use different context size (2, 3, 4 and 5) Use small and large bilingual glossary Compared it with the seed words algorithm Compared with traditional orthographic similarity measures: LCSR and MEDR


Experiments : Experiments BASELINE: random MEDR: minimum edit distance ratio LCSR: longest common subsequence ration SEED: the "seed words" algorithm WEB3: the Web-based similarity algorithm with the default parameters: context size = 3, small glossary, stop words filtering, no lemmatization, no reverse context lookup, no TF.IDF-weighting NO-STOP: WEB3 without stop words removal WEB1, WEB2, WEB4 and WEB5: WEB3 with context size of 1, 2, 4 and 5 LEMMA: WEB3 with lemmatization HUGEDICT: WEB3 with the huge glossary REVERSE: the "reverse context lookup" algorithm COMBINED: WEB3 + lemmatization + huge glossary + reverse context lookup


Resources : Resources We used the following resources: Bilingual Bulgarian / Russian glossary: 3 794 pairs of translation words Huge bilingual glossary: 59 583 word pairs A list of 599 Bulgarian stop words A list of 508 Russian stop words Bulgarian lemma dictionary: 1 000 000 wordforms and 70 000 lemmata Russian lemma dictionary: 1 500 000 wordforms and 100 000 lemmata


Evaluation : Evaluation We order the pairs of words from the testing dataset by the calculated similarity False friends are expected to appear on the top and the cognates on the bottom We evaluate the 11pt average precision of the obtained ordering


Results (11pt Average Precision) : Results (11pt Average Precision) Comparing BASELINE, LCSR, MEDR, SEED and WEB3 algorithms


Results (11pt Average Precision) : Results (11pt Average Precision) Comparing different context sizes; keeping the stop words


Results (11pt Average Precision) : Results (11pt Average Precision) Comparing different improvements of the WEB3 algorithm


Results (Precision-Recall Graph) : Results (Precision-Recall Graph) Comparing the recall-precision graphs of evaluated algorithms


Results: The Ordering for WEB3 : Results: The Ordering for WEB3


Discussion : Discussion Our approach is original because: Introduces semantic similarity measure Not orthographic or phonetic Uses the Web as a corpus Does not rely on any preexisting corpora Uses reverse-context lookup Significant improvement in quality Is applied to original problem Classification of almost identically spelled true/false friends


Discussion : Discussion Very good accuracy: over 95% It is not 100% accurate Typical mistakes are synonyms, hyponyms, words influenced by cultural, historical and geographical differences The Web as a corpus introduces noise Google returns the first 1 000 results only Google ranks higher news portals, travel agencies and retail sites than books, articles and forums posts Local context could contains noise


Conclusion and Future Work : Conclusion and Future Work Conclusion Algorithm that can distinguish between cognates and false friends Analyzes words local contexts, using the Web as a corpus Future Work Better glossaries Automatic augmenting the glossary Different language pairs


Questions? : Questions? Cognate or False Friend? Ask the Web!