tgt
stringlengths
1
2.76k
src
stringlengths
1
3.03k
lp
stringclasses
15 values
(אדם2), אוקיי, אני אעשה את זה.
(PERSON2) Okay, I'll do that.
en-he
(אדם11), כן, תודה.
(PERSON11) Yeah, thanks.
en-he
אני עכשיו עובד על, או בעצם מקבל עדכון, על הגשת מסמך הדמו, שצריך להגיש ביום שישי הבא, אם אני לא טועה.
I'm now working on, or getting to update, the demo paper submission, which is due next Friday, if I'm not mistaken.
en-he
ואז יש לנו כמה דוחות פרויקטים לעשות ואני מתחיל לאסוף מסמכים – כמו, או קישורים למסמכים, לא המסמכים עצמם עדיין, לשני סשנים.
And then I have some project reports to do and I'm starting to gather documents - like, or links to the documents, not the documents themselves yet, for the two sessions.
en-he
(אדם11) <שיחה_מקבילה> גם עבור ASR עלינו להוסיף מטרי משלים –
(PERSON11) <parallel_talk> For ASR as well we should add a complimentary metric -
en-he
למעשה מעבר לכך, עלינו להוסיף מטרי משלים – המטרי המשלים הזה. </שיחה_מקבילה>
Actually more, we should add complementary met- these complementary metrics. </parallel_talk>
en-he
אז הזכרת אחד שהוא הערכת מסנן ניבולי פה.
So, you mentioned one and that is profanity filtering m evaluation.
en-he
אז <שיחה_מקבילה> יש רשימה ברורה של דברים שאסור שיופיעו בפלט </שיחה_מקבילה>, נכון?
So <parallel_talk> have an explicit list of things which must not appear in the output </parallel_talk>, right?
en-he
<שיחה _מקבילה> והערכה של מילים נדירות.
<parallel_talk> And then rare words evaluation.
en-he
וכאן יש רשימה ברורה של שמות ומונלים שאנחנו לא רוצים שיופיעו בפלט, עם ניקוד לא רק על שיעור הסדר </שיחה_מקבילה> אלא על ידי משהו שמבסס את הדברים האלה, נכון?
And here have an explicit list of names and terms that we do want to appear in the output, scored not simply by (order) rate </parallel_talk> but scored by something which directly accounts these things, right?
en-he
כן, אני מסכים לגמרי.
Yes, so I totally agree.
en-he
זה דורש הכנה ידנית של ההפניה.
This involves also manual preparation of the reference.
en-he
אז, <שיחה_מקבילה> זקוק להפניות שמוכנות באופן ידני. </שיחה_מקבילה>
So, <parallel_talk> both need manually created references. </parallel_talk>
en-he
יש לי הערה מ-[אדם1].
I have one remark from [PERSON1].
en-he
אז <שיחה_מקבילה> [אדם1] הקים מילון תרגום.
So <parallel_talk> [PERSON1] has compiled some translation dictionary.
en-he
והיא תשתף אותו בקרוב. </שיחה_מקבילה>
And she should share it soon. </parallel_talk>
en-he
אז זה יכול להפוך לחלק ריק של הדבר השני של המילה הנדירה –
So, this could become an empty part of the second thing of the rare word -
en-he
אז <שיחה_מקבילה> זו תהיה הערכה של מילים נדירות עבור תרגום מכונה ואז מ-[פרויקט210] אנחנו עשויים לקבל מילון כלשהו. </שיחה_מקבילה>
So <parallel_talk> this would be the rare word evaluation for MT and then from [PROJEC210] we are also likely to get some dictionary. </parallel_talk>
en-he
אז זה יהיה הדומיין של [פרויקט210] מן הסתם.
So that would be the [PROJEC210] domain obviously.
en-he
ו-[אדם2], אפשר לבקש ממך לעבור על דומיין <לא מובן/> וליצור כזה מילון על ידי מעבר על הפלטים ומה שחסר בהם.
And [PERSON2], could I ask you to revise the <unintelligible/> domain and create such a dictionary by looking at the outputs and what is missing in the outputs.
en-he
כמו רשימה קצרה – רשימה של מילים שאנחנו אוהבים ולא אוהבים בדומיין <לא מובן/>.
So, like shortlisting - listing the words that we do like, and we do not like in the <unintelligible/> domain.
en-he
אפשר לעשות את זה?
Could you do that?
en-he
(אדם11) נראה שאפשר לעשות את זה אז בואו ננסה.
(PERSON11) It should be doable, so let's give it a try.
en-he
זה יהיה נהדר אם אפשר יהיה לעשות את זה.
It would be like - it would be great if you managed to do that.
en-he
אז אני אכתוב את זה.
So, I'll write it down.
en-he
אז <שיחה_מקבילה> מכוונים כלפי המועד האחרון של [ארגון62] עם מאמר על מחקר כתוביות. </שיחה מקבילה>
So, <parallel_talk> aiming towards [ORGANISATION62] deadline with subtitler study paper. </parallel_talk>
en-he
כן <שיחה_מקבילה> ואז מקורות מרובים עם [אדם7] או על בסיס [אדם7] או – </שיחה_מקבילה>
Yes <parallel_talk> and then then a multi-source with [PERSON7] or based upon [PERSON7] or - </parallel_talk>
en-he
ועבור מזהה השפה, אני סקרן לגבי האופן שבו תוכל לשלב את זה בדיוק כי זה דורש מאיתנו לשקול מקורות ASR מרובים, ערוצים מרובים, אז מה יהיה מקרה שימוש עבור מזהה השפה?
And for the language ID, I'm curious how do you want to integrate it exactly because it already involves considering multiple ASR sources, multiple channels, so what would be the use case for the for the language ID?
en-he
איך אנחנו מחברים את זה?
How do we plug that in?
en-he
(אדם13) זה יהיה עובד אודיו לטקסט והוא יפלוט חותמות זמן לחלון זמן של 2 או 4 שניות והקטגוריה.
(PERSON13) It will be audio to text worker, and it will emit time stamps like for four or two second window and the class.
en-he
<שיחה_מקבילה/> שקט, צ'כית, גרמנית, אנגלית, ואז אני אודיע לאחרים איך משתמשים בזה בצינור.
<unintelligible/> silence, Czech, German, English, and then I will let on the others how they want to use it in the pipeline.
en-he
(אדם11) זה חשוב.
(PERSON11) That is important.
en-he
כתוב את זה כאן בסיכום של <לא מובן/> כי זה עבור [אדם2].
Please write this down here into the documenter this <unintelligible/> summary, because that is for [PERSON2].
en-he
וגם [אדם2], כשאתה מדבר עם [אדם10], וגם עם [אדם5].
And also [PERSON2], when you talk to [PERSON10], well and also possibly [PERSON5].
en-he
אם יש לנו את הכלי הזה, אנחנו צריכים (לשלוח) את האודיו ל-ASR מרובים או עובדים מרובים <לא מובן/>
If we have this tool, we need to (ship) the audio to multiple ASR or multiple workers <unintelligible/>
en-he
אנחנו נצטרך (לשלוח) בנפרד את האודיו ל-ASR באנגלית, ל-ASR בגרמנית ול-ASR בצ'כית, לדוגמה תלוי ב-<רעש_אחר/>
We will separately need to (ship) the audio to the English ASR, separately to the German ASR and the Czech ASR, for example depending on the <other_noise/>
en-he
וגם לעובד עם מזהה השפה ואז נצטרך למזג את הפלטים ואת הכלי שאותו אין לנו עדיין.
And also, to this language ID worker and then we need to merge these outputs, and this is tool which we do not have yet.
en-he
זה הכלי מרובה המקורות, שיצפה בפלטים של הטקסט ויעקוב אחר חותמות הזמן והוא גם יפלוט את – הוא גם כנראה ייצור את הפלט אל תוך שלושה ערוצי שפה לשימוש בשקט או ב-ASR הנכון.
That's the multi-source tool, which will be observing the text outputs and also following the time stamps and it would emit the - it would probably like produce the output into three separate language channels for using either silence or the correct ASR.
en-he
לכן, אנחנו צריכים מסנן עבור ה-ASR כך שה-ASR יהיה שקט אם זו השפה הלא נכונה והוא הטקסט המזוהה אם זו השפה הנכונה.
So, we need a filter for the ASRs, so that the ASR is silent if it's the wrong language and it is the recognized text if it's the correct language.
en-he
אני פשוט המצאתי את הסוג הזה של ההגדרה.
I just made it up this type of setup.
en-he
הגדרה נוספת תהיה אותו צליל שנשלח ל-ASR ולבדוק השפה הזה.
Another of setup would be that same sound is shipped to ASR and this language checker.
en-he
בודק השפה הזה הוא בעצם חלק מה-ASR, והוא משתיק את ה-ASR אם זו השפה הלא נכונה.
This language checker is essentially a part of the ASR, silencing the ASR if it's the wrong language.
en-he
זו גם אפשרות.
That's also an option.
en-he
אנחנו צריכים להבין מהי הדרך הטובה ביותר לשלב את מזהה השפה מבחינת המטרות שלנו.
So, we need to figure out which way of integrating the language ID is the best for our purposes.
en-he
אז המשיכו לחשוב על זה, מה הם הצינורות שלנו.
So, please keep thinking about this like what are our pipelines.
en-he
באופן אידיאלי, אני חושב שיהיה כמה שפחות בלבול ניהולי אם המודלים של תרגום המכונה שלנו יהיו רב לשוניים.
Ideally, I think that the least like management confusion and so on would arise if our MT models were multilingual.
en-he
אם הם יתמכו בשפות ממקורות שונים ויתרגמו מכל אחת מהשפות האלה לאנגלית.
If they supported different source languages and they were translating from any of these languages into English.
en-he
אולי כדאי לעשות רק עותק, אם האנגלית משמשת כקלט, כך שיהיה לו ASR רב לשוני בהתחלה.
Possibly doing just a copy, if the English was given as the input, so we would have multi-lingual multiple ASRs at the beginning.
en-he
הכל יהיה לאנגלית ואז מאנגלית זה יילך לכל השפות.
All going into English, and then from the English will go into all of the languages.
en-he
ומאוחר יותר כשל-[אדם12] יהיה מודל רב לשוני של ASR, לא יהיה צורך יותר במזהה שפה.
And later when [PERSON12] would have multi-lingual model of ASR, that language ID would not even be needed.
en-he
(אדם3) העניין הוא, ככל שנבצע את הניתוח בצורה נכונה יותר, עניין אחד הוא ההערכה המעמתת האפשרית, רק כדי לציין שהאילוצים בעצם משפיעים על הפלט עצמו.
(PERSON3) The thing is, as long as we do the analysis more properly, one thing is the possible adversarial evaluation, just to indicate that that the constraints are actually having an impact on the actual output.
en-he
זה יהיה נחמד וכן כנראה ניתוח תשומת הלב.
That that would be nice and yeah possibly the attention analysis.
en-he
אבל אני לא בטוח שזה ייתן לנו את אותן תשובות, למעשה, אם המודל קשוב או לא קשוב לאילוצים.
But I'm not sure it might give us the same answers, basically, if the if the model attends or does not attend to the constraint.
en-he
(אדם4) הייתי חייב לבדוק כמה דוגמאות של תשומת לב, לא ביצעתי סטטיסטיקה כלשהי, והוא ראה את האילוצים וכשהוא תרגם את האילוצים, נראה שהאילוצים היו נתונים, ואני חושב שזה יהיה יעיל אם המערכת הייתה עושה טעויות אבל למעשה היא לא עשתה טעויות בעניין הזה.
(PERSON4) I had to look just at a few examples of attention, I did not do any statistics or anything, and it looks at the constraints, and when it translates the constraints, it looks like the constraints given, and I think it would be useful if the system made any errors, but it in fact does not do any in that sense.
en-he
היא עשתה כמה טעויות תרגום כלליות אבל זו לא ה(תופעה) שאנחנו מנסים לפתור.
It makes some general translation errors but not the (phenomena) we are trying to solve.
en-he
בחזרה להערכה האוטומטית, הבעיה עם חוסר ההתאמה היא שהפלט מוטה נכון אבל ההקשר שונה.
Well, going back to the automatic evaluation, the problem with mismatch is that the output is correctly inflected, but the context is different.
en-he
לכן, זה לא אותה צורת מילה כמו בהפניה, נכון?
Therefore, it is not the same word form as in reference, right?
en-he
(אדם4), כן, אבל אני גם בדקתי אם ההקשרים הם תרגומים תקפים וברוב המקרים, הם כן.
(PERSON4) Yeah, but I have also checked if the contexts are valid translations and in most cases they are.
en-he
כמו ב-100 הדוגמאות הראשונות שסומנו כשגיאה על ידי הערכה אוטומטית 91 מתוכן היו מוטים נכון בהקשרים נכונים.
Like in the first 100 examples that were marked as error by automatic evaluation 91 of them were correctly inflected in correct contexts.
en-he
(אדם4) ואני חושב שרק שניים או שלושה משפטים היו מוטים נכון בהקשר לא נכון, כאילו התרגום היה שגוי.
(PERSON4) And I think only two or three sentences were correctly inflected in the incorrect context, like the translation was wrong.
en-he
ויש גם מקרים שבהם התרגום היה שגוי לחלוטין מפני שהמשפט היה שגוי לחלוטין והחלק חייב היה להיות <לא מובן/> אבל זה היה רק באחד או בשני מקרים.
And then there are some cases where the translation was totally wrong, because the sentence was really wrong, and the part must be <unintelligible/>, but that was just like one or two cases.
en-he
(אדם3) אבל אתה מנסה למזג את הצורות שונות כשאתה מבצע הערכה, נכון?
(PERSON3) But you are trying to match the divert forms when you are evaluating it, right?
en-he
אין לך -
You do not do any-
en-he
(אדם 3) – למטיזציה של ההפניה וה-
(PERSON3) -lemmatization of the reference and the-
en-he
(אדם4) גם, גם, אני משווה את גם את צורת פני השטח והלמטיזציה.
(PERSON4) Both, both, both, I'm matching both the surface forms and lemma.
en-he
(אדם3) אני מבין, כן, אלה הן שתי הנקודות, כן?
(PERSON3) Oh I see, oh yeah, those are the two scores, right?
en-he
(אדם4) זו בדיוק הדילמה, הניקוד של ה-(אירופאי?) הוא לא הטבלה כי הטבלה לא תתאים למסמך אז <unintelligible/>
(PERSON4) It is just the dilemma score for the (European?) is not in the table because of the table like won't fit into the paper so <unintelligible/>
en-he
הכיסוי הוא תמיד 97 אחוז.
Coverage is always like 97 percent.
en-he
הוא פשוט מייצר את הלמטזציות הנכונות, רק –
It just generates the correct lemmas, just-
en-he
(אדם3) כן, אוקיי, הבנתי, הבנתי.
(PERSON3) Oh yeah, okay, I get it, I get it.
en-he
אז למעשה, כיסוי הלמטיזציה אומר אם האילוץ קיים, וההבדל של הכיסוי של פני השטח מציין שזה יכול להיות מוטה בצורה שגויה אבל זה לא המקרה.
So basically, the lemma coverage says whether the constraint is there, and the surface coverage difference suggests that it might be incorrectly inflected, but that is not the case.
en-he
(אדם3) כן, אני רק לא יודע אם אתה תופס <בלתי מובן/> כשדיברנו על [אדם1],
(PERSON3) Oh yeah, just I do not know if you if you, if you catch <unintelligible/> when we were talking about [PERSON1],
en-he
אבל יש דבר אחד, שה(פגיעה) טובה יותר אבל דבר אחר הוא שהלמיטיזציות מוטות בצורה נכונה יותר באופן ידני,
but one thing, is, that, that the (blow) is better but the other thing is that the lemmas are actually properly inflected manually,
en-he
אבל זה משהו אחר. אין בעיה, זה טוב בעצם <בלתי מובן/> כן.
but that is that is another, no, no problem, it is actually good that is <unintelligible/> right, yeah.
en-he
אז אני חושב שיכול להיות.
So, I was thinking there might be.
en-he
אבל אני לא בטוח שיש לנו נתוני בדיקה של זה אבל יכול להיות שננסה להפעיל את המודלים עם סגנון מסוים של מעבר?
But I'm not sure whether we have some test data for that, but we might try to play around with the models with some sort of style transfer?
en-he
כמו שאתם יודעים, אנחנו יכולים להשתמש באילוץ ולנסות להשתמש באילוצים של מילים נרדפות במקום ולראות או להשוות את ההבדל בין הפלטים.
As you know, we can use a constraint and try to use a synonymous constraint instead and see or compare how, how, what is the difference between the outputs.
en-he
אבל זה כמו שאני מתאר את זה בעמימות מפני שאני בעצמי לא ממש יודע איך לעשות את זה אבל זה יכול להיות תרחיש שונה ומעניין.
But this is, like I'm just describing it vaguely because I myself do not have an exact idea how to do it, but it might have been an interesting, distinct scenario.
en-he
אני לא יודע, [אדם1], אתה יודע על נתוני מעבר סגנון עבור אנגלית צ'כית? יש לנו משהו?
I do not know, [PERSON1], do you know about any style transfer data for English Czech? Do we have something?
en-he
(אדם4) <בלתי מובן/>
(PERSON4) <unintelligible/>
en-he
(אדם1) כשאתה מדבר על העברת סגנון, דוסן, אתה מתכוון שיש משהו בשפה הכתובה, ואתה בצד של היעד, שיהיה בשפה המדוברת או משהו כזה?
(PERSON1) By style transfer, Dusan, you mean that there would be something which is in the written language, and you would be the target side, would be in the spoken language or something like that?
en-he
(אדם3) כן, יש לי מידע מוגבל על המשימה הזו.
(PERSON3) Well, that is the thing like, I have only a limited knowledge about the task.
en-he
אבל אני יכול לתאר שאתה רוצה לשכתב משפט שלא נכתב על ידי גבר, אבל נכתב על ידי אשה במקום זאת, או, אני לא יודע, יכול להיות לך הסבר מדעי או לא מדעי על תופעה מסוימת.
But I can imagine that you want to rewrite a sentence that it is not written by male, but it is written by a female instead or, I do not know, like you can have a scientific and unscientific explanation of certain phenomena.
en-he
אני לא בדיוק מכיר את הסטים של הנתונים ומה הם מכסים.
I'm not really that familiar with the exact data sets and what they cover.
en-he
והרעיון הוא שהסגנון של המשפטים מוגדר בצורה די עמומה.
And the idea is that the style of the sentences gets quite vaguely defined.
en-he
אז שוב, לא כל כך בטוח.
So again, not so sure.
en-he
(אדם1) יש לנו קורפוס של טרנספורמציות של משפטים.
(PERSON1) We have this corpus of sentence transformations.
en-he
ודבר אחד יכול להיות מוגדר כהעברת סגנון, לדוגמה, להפוך את המשפטים ליותר כלליים.
And one thing that is there and could be labelled as style transfer is for example, making the sentences more general.
en-he
אז פרטים נשמטים, המשפט הופך להיות פשוט יותר ואז הוא נשמע דיבורי.
So, details are omitted, the sentence is simplified, and then making the sentence sound colloquial.
en-he
כן, רק ניסוח מחדש.
Well, just a paraphrase.
en-he
ואני לא בטוח אם יש לנו משהו כמו קיצור המשפט.
And I'm not sure if we have anything like shortening of the sentence.
en-he
בעניין קיצור, מתאוס מבצע ניסוי מסוים עם אנגלית-צ'כית, אבל אין לנו עדיין נתונים של הפניות.
For shortening, Matous is doing some experiment with English–Czech, but we do not have any reference data there.
en-he
אז אין לנו מערך נתונים של משפטים דחוסים.
So, we do not have any sentence compression dataset.
en-he
אז אולי ההכללה יכולה להיות מעניינת אבל אני חושש שיש יותר מדי הכללות טובות, אז ההכללה של ההפניה היחידה יכולה להיות מוגבלת מדי מכדי לציין משהו על איכות ההכללה שלך.
So maybe the generalization could be of interest but I'm afraid that there are too many different good generalizations, so the single reference one would be too limited to tell you anything about the quality of your generalization.
en-he
ואני לא מודע לטרנספורמציות של מגדר, לדוגמה, אפשר לעשות את זה <בלתי מובן/> לצ'כית, אז אולי האדם נכון יהיה רודולף רוזה לשאול אותו אם הוא אי פעם יצר מערך נתונים כזה.
And I'm not aware of any like gender transformations, for example, this could be done <unintelligible/> for Czech, so maybe the right person would be Rudolf Rosa to ask him if he has ever generated any such dataset.
en-he
אז יהיה ייצור מבוסס-שורש של מקבילים מסוימים של משפטים.
So, there will be a root-based generation of some sentence counterpart.
en-he
(אדם3) אני מניח שזו שאלה מעניינת אם אנחנו יכולים להשתמש באילוצים כדי לאכוף את המגדר של הדובר בתרגום.
(PERSON3) I guess that is actually like an interesting question whether we can use the constraints to enforce this kind of the like gender of the speaker in the in the translation.
en-he
(אדם1) זה בעצם רעיון טוב מאוד.
(PERSON1) That is actually a very good idea.
en-he
אז אנחנו יכולים להתמקד וליצור תת-חלק של [פרויקט1] לסט בדיקה שיכסה את זה.
So, we could focus on that and create a particular sub-part of [PROJECT1] test set that would cover that.
en-he
כך שלפעמים אנחנו יודעים את המגדר של הדובר אז אם אתה... אולי אני אשתף את המסך ונעיין בזה.
Ah so that we sometimes we know the gender of the speaker so if you are... maybe I'll share the screen and browse that.
en-he