detect language before NER

[schloett]

The NER-service supports NER for English and German (if no language definition is there, it assumes English). Detect the language client-side and inform the NER-service about it. To request NER for a specific language, the format looks like:

{
    "paragraphs" : [<the paragraphs>],
    "language" : "de"
}

However, the quality for German NER seems to be far below the English one.

[schloett]

language is detected in v4.3.0

[chseifert]

What languages are detected? Users observed to receive Dutch results (Rijksmuseum), while on a German page. Is there only "de" "en" and "others"?
Re-opened, but might just have been for discussion.

[schloett]

Unfortunately, the piece of information relevant to this issue is missing:
What is the language of the extracted keywords? (i.e. the query).
The language detection in C4 is used to select the strategy for keyword extraction (atm calling the NER-service for English and applying some heuristic for all other languages). Therefore, the extracted keywords must be in English (in case the language has been detected as English, then the keywords returned from the NER-service are English) or in the language of the page. Otherwise we would have invented a magic heuristic, which translates German to Dutch.
Imho, the returned Dutch results are beyond the scope of C4 and in the responsibility of the federated recommender.
For the sake of completeness, the list of possibly detectable languages:
"ab": "Abkhazian",
"af": "Afrikaans",
"ar": "Arabic",
"az": "Azeri",
"be": "Belarusian",
"bg": "Bulgarian",
"bn": "Bengali",
"bo": "Tibetan",
"br": "Breton",
"ca": "Catalan",
"ceb": "Cebuano",
"cs": "Czech",
"cy": "Welsh",
"da": "Danish",
"de": "German",
"el": "Greek",
"en": "English",
"eo": "Esperanto",
"es": "Spanish",
"et": "Estonian",
"eu": "Basque",
"fa": "Farsi",
"fi": "Finnish",
"fo": "Faroese",
"fr": "French",
"fy": "Frisian",
"gd": "Scots Gaelic",
"gl": "Galician",
"gu": "Gujarati",
"ha": "Hausa",
"haw": "Hawaiian",
"he": "Hebrew",
"hi": "Hindi",
"hmn": "Pahawh Hmong",
"hr": "Croatian",
"hu": "Hungarian",
"hy": "Armenian",
"id": "Indonesian",
"is": "Icelandic",
"it": "Italian",
"ja": "Japanese",
"ka": "Georgian",
"kk": "Kazakh",
"km": "Cambodian",
"ko": "Korean",
"ku": "Kurdish",
"ky": "Kyrgyz",
"la": "Latin",
"lt": "Lithuanian",
"lv": "Latvian",
"mg": "Malagasy",
"mk": "Macedonian",
"ml": "Malayalam",
"mn": "Mongolian",
"mr": "Marathi",
"ms": "Malay",
"nd": "Ndebele",
"ne": "Nepali",
"nl": "Dutch",
"nn": "Nynorsk",
"no": "Norwegian",
"nso": "Sepedi",
"pa": "Punjabi",
"pl": "Polish",
"ps": "Pashto",
"pt": "Portuguese",
"pt-PT": "Portuguese (Portugal)",
"pt-BR": "Portuguese (Brazil)",
"ro": "Romanian",
"ru": "Russian",
"sa": "Sanskrit",
"bs": "Serbo-Croatian",
"sk": "Slovak",
"sl": "Slovene",
"so": "Somali",
"sq": "Albanian",
"sr": "Serbian",
"sv": "Swedish",
"sw": "Swahili",
"ta": "Tamil",
"te": "Telugu",
"th": "Thai",
"tl": "Tagalog",
"tlh": "Klingon",
"tn": "Setswana",
"tr": "Turkish",
"ts": "Tsonga",
"tw": "Twi",
"uk": "Ukrainian",
"ur": "Urdu",
"uz": "Uzbek",
"ve": "Venda",
"vi": "Vietnamese",
"xh": "Xhosa",
"zh": "Chinese",
"zh-TW": "Traditional Chinese (Taiwan)",
"zu": "Zulu"

[chseifert]

Just to be completely sure: The language is also sent in the user profile (or just used to parametrize the keyword selection)?

[schloett]

The language is only used to parametrize the keyword selection. Languages in the user profile are only sent, if the user specifies and discloses languages in the user profile settings.

[chseifert]

Hmm, this seems to be annoying for users - not to have the content mapped automatically to the web page. I add Thomas to discuss the privacy issues for sending the detected language. Programmatically this should not be a big issue as I understand.

[ThomasCerq]

I'm not sure I understand the issue...

As far as I understand:

• The language is extracted from the context (the web page being seen) and used to build a query in this languages. In this case, I don't see privacy issues, looking at the content the user is looking at is the only way to give recommendations.
• The languages understood by the user are specified in the user profile. The user can decide to share this information or not. In case, it is shared, it is sent with the profile to the FR so it can recommend resources in the languages understood by the user. In this case, the user privacy is addressed by the disclose/hide information button in the UI.

I guess we could try to map both information: if the web page is in english, then the recommendations should be in english too. But it would also make sense to me if recommendations in spanish were sent (if the user explicitly said she understands spanish).

Does it make sense?

[ThomasCerq]

Just FYI, in the user profile UI, only european languages are included. But happy to change it if needed.

[chseifert]

Thanks for you answer.

Yes, it makes sense. Language in User Profile means "the languages I speak and understand" - The language we detect is the language of the page (also presumably "a language I understand").
Both can be used for source selection. My only concern is that

• if the user does not specify a language (a likely use case) than the results will be strange
• if we would include both information (manual setting and automatic detection), it gets complicated
• if we only include the automatic detection results will be only in one language (that of the web page), recall might be low.

As for the process of "keywords in a specific language" to "results in the specific language" without an explicit language flag sent to the recommender - I don't think this works. What they have implemented (according to deliverable D3.3.) is the source selection based on language when they receive an explicit flag.

I am not sure how to solve this.

[schloett]

From my point of view, it is an issue of the recommender to provide the "correct" results. On clientside, we could only ease the burden of the recommender.
"Correct" results would be either in one of the languages disclosed in the user profile (if any) or the language of the page. If the user discloses the languages of the profile, we send them anyway. I agree with you, that it is unlikely, that a user provides this information, hence what remains is the language of the page. This information is already encoded in the query and the recommender might detect the language of the query and filter / rank the results accordingly. When such a mechanism is in place at the recommender, I am perfectly fine with providing language information about the current page, in order to overcome the "detect language of the query" step. However, as long as there is no such mechanism is in place, I would not change the profile again to send information which is never used.
Furthermore, to provide correct results, imho there are other modifications necessary between federated recommender and partner recommenders:

In this example, the language is unknown for all items from DDB and Mendeley, while the language of all items from ZBW is German, which is obviously not correct for the first result.

[chseifert]

Jörg, you are right, with a) we can only ease the burden of the recommender and b) should not employ API changes unless we are sure the information will be used.
I am just taking up user comments and trying to find out i) whether and ii) how we could tackle them.

Following the discussion I come to the conclusion that

• we should use the automatically detected language to generate a good query (as is)
• let the user input his/her spoken languages manually (as is)

Potential (simple) improvements would be

• in the profile settings rename the paragraph "Languages" to "Spoken Languages" and/or provide a tooltip saying "The language information is used to adapt your search results. Results corresponding to your spoken language will be ranked higher in the list of search results"

[ThomasCerq]

Following the discussion I come to the conclusion that

• we should use the automatically detected language to generate a good query (as is)

  Agreed.

• let the user input his/her spoken languages manually (as is)

  Agreed too. But you're right when you say it's very unlikely that users will enters languages in the UI. Could we have a kinda of screen tip when there are several languages in the results? It would say something like "Results are in english, german and spanish. If you don't understand some of them, you can set the languages you understand in the user profile UI". (Should be displayed only the first time or from time to time.)

Potential (simple) improvements would be

• in the profile settings rename the paragraph "Languages" to "Spoken Languages" and/or provide a tooltip saying "The language information is used to adapt your search results. Results corresponding to your spoken language will be ranked higher in the list of search results"

  That's a good idea, and it will be included in the next version.