Leveraging Loanword Constraints for Improving Machine Translation in Low-resource Settings

What this episode is about

The seminar focuses on leveraging loanword constraints to enhance machine translation systems for low-resource languages, with a specific emphasis on the Emakua language. Presenter Felerino Ali outlines methods to incorporate bilingual dictionaries and loanword mappings into training processes, addressing challenges like data scarcity and vocabulary gaps in neural machine translation.

Key takeaways

• Machine translation is evolving, but low-resource languages still face a data desert.
• Loan words and code-switching could be the secret sauce for translating complex African languages.
• Over 7,000 living languages exist, yet less than 200 are covered by current translation tools.
• Neural networks are revolutionizing translation, but they still stumble over vocabulary gaps in underrepresented tongues.
• Language diversity is a goldmine for NLP; tapping into it can bridge global communication divides.

Who speaks in this episode?

• Stephan (Host) · 5 min speaking
• Felerino Ali (Guest) · 38 min speaking
• Steve (Host) · 1 min speaking
• Millison (Guest) · 1 min speaking
• Germa (Guest) · 1 min speaking

What is discussed, and when

1. 0:00 – 5:00: Emacua, Felerino Ali, NLP, Microsoft Research Africa
2. 5:00 – 10:00: Mozambi, Emacua
3. 15:00 – 20:00: LLM

Notable quotes from "Leveraging Loanword Constraints for Improving Machine Translation in Low-resource Settings"

• Felino is a posttock at Microsoft Research Africa here in Nairobi and a PhD candidate at the University of Porto in Portugal. his research focuses on NLP processing.

  about Felerino Ali · at 0:18

• Gaps through structured lexical integration and I believe Phil presented this at EMNLP earlier this month. So, we're super excited to have him. And over to you, Famino.

  about EMNLP · at 1:13

• As Stephan mentioned, this is basically a paper that recently have, presented at PMLP, which is this one.

  about PMLP · at 2:03

• Continent we can see that there is also other more underrepresented language which is the case of like for example is opening countries which if you see in this map the they…

  about Mozambi · at 7:07

• Will explain later. Then nowadays there's a very interesting trend which is using LLM following and of course these models are being trained to like set certain checkpoint and then they you can use rag to you know give them more information.

  about LLM · at 14:55

Who and what is mentioned in this episode?

People

• Felerino Ali — Person · Main speaker · 1 mention · first at 0:05
  Seminar presenter describing research on lexicon-guided neural machine translation for low-resource settings, focusing on loanword constraints and dictionary integration for Emakhuwa–Portuguese MT.

Companies & orgs

• Microsoft Research Africa — Company · Affiliation · 1 mention · first at 0:18
  Employer of the presenter, contributing to the research.
  Felino is a posttock at Microsoft Research Africa here in Nairobi and a PhD candidate at the University of Porto in Portugal. his research focuses on NLP processing.

Concepts & topics

• Emacua — Concept · Target language · 4 mentions · first at 1:00
  Low resource African language focused on in the research.
  Integrate bilingual dictionaries and loan word mappings into training process to address these challenges for emacua I hope I pronounced that right which is a low resource band…
• NLP — Concept · Research field · 1 mention · first at 0:28
  Natural Language Processing is the broader field of the research discussed.
  PhD candidate at the University of Porto in Portugal. his research focuses on NLP processing particularly for low resource African languages with an emphasis on developing…
• LLM — Concept · Technology used · 1 mention · first at 15:00
  Large Language Models used in the research for improving translations.

Places & events

• Mozambi — Place · Research setting · 3 mentions · first at 7:12
  Country where the target language Emacua is spoken.
• EMNLP — Event · Presentation venue · 1 mention · first at 1:15
  Conference where the research was presented.
• PMLP — Event · Presentation venue · 1 mention · first at 2:11
  Conference where the research was presented.

Full transcript with timestamps

0:02 I will take that as a yes. Okay. Well, welcome to today's seminar. Uh, today we have Felerino Ali presenting and he'll be talking about leveraging loan word constraints for improving machine translation in low resource settings. Felino is a posttock at Microsoft Research Africa here in Nairobi and a PhD candidate at the University of Porto in Portugal. uh his research his research focuses on NLP processing particularly for low resource African languages with an emphasis on developing neural machine translation systems and improving methods for evaluating machine translation performance in these settings. Um and today he'll be uh talking about how you can translate between worlds when data is scarce and the language is full of twists. So thinking about loan words, code switching and orthographic quirks. Um he'll discuss how to apply lexicon guided neural machine translation integrate bilingual dictionaries and

0:58 loan word mappings into training process to address these challenges for emacua I hope I pronounced that right which is a low resource band to language. Um this method suggests a promising direction for low resource m machine translation by bridging data scarcity and vocabulary gaps through structured lexical integration and I believe uh Phil presented this at EMNLP earlier this month. So, we're super excited to have him. And over to you, Famino.

1:32 Thank you. Thank you. And first of all, happy to be here. And um can you please confirm if you can see my screen? Yes, go ahead. Let me know.

2:01 So, um, as Stephan mentioned, this is basically a paper that recently have, uh, presented at PMLP, uh, which is this one. this one and uh just to make the presentation more easy to follow I will start back on the like background of machine translation and then I'll go through the paper and what was the idea behind it. So going back for like the context of of the whole paper which is machine translation which is the task of converting the meaning from one language into another by means of a machine which we can see in this slide where we are trying to translate from Portuguese into a mua the sentence like uh do you can you speak and read a map? So the translation is something that is there.

3:05 So this is a machine translation task which is something that had been around for a long time and early machine translation system uh dated like in 1950s and the first machine translation system at the at the time were based on rules like rule based system. It used to have like um a lot of insights from linguistics. So this systems basically uh it was like um something that we had in the in the right side right side of this slide where you want to translate the word and then depending of for example the subject and object you could have like a different translation. So linguistics uh could come up and design these rules and constructions. But then as you can imagine this was was very difficult to maintain. You have to go back and and refine when you know language is so complex and you can you cannot predict all the possibilities and rules that can can define a language. So that's why that time it was

4:21 a shift from the rule based system to like datadri approach in which you learn from data. If you have a lot of data you can you know train the models in the statistical way which was the the first type of machine translation and nowadays is like using neural machine translation which are based on neural networks. uh and of course during the time many different approach are created from like recurrent neural network up to now that we have uh transformer based uh neural machine translation. So um one problem with the a datadriven approach is the data the data limit this this approach and nowadays we have this um gap between so-called high resource and low resource languages where the high resource language are languages like English that have like a wide presence on the internet you have enough data to train the models but then in the tail of this graph you can see that there a language that have really

5:30 low resource which is the case of many African languages and how can we train the models uh to also you know fill this gap between the higher sourcing languages like the problem that we are uh dealing in this area of research. So um you can expect many of the tools that we have now in the market on translation and you see you see in terms of uh quality of quality in terms of translation translating like African languages or low source languages are still are still uh poor and also the coverage of these systems are also poor.

6:16 So less than I would say less than 200 languages are covered by these existing tools which is very low compared to like living living languages in the world like more than 7,000 and that's why we have like nowadays a very active research um the community which is trying to tackle this problem and for example in the African context many uh researchers come come to tackle this problem. And what as we can see here in this map, the heat map means that the countries that we mostly have contributed to research on the machine translation uh research.

7:02 So but even if we see on this if you look at this context of African continent we can see that there is also other more underrepresented language which is the case of like for example is opening countries which if you see in this map the they are less content to to the body of knowledge and that's why I actually this was the motivation for me PhD and Um and my focus was contributing with the one of the languages spoken this opening countries in in this case Mozambi which is a language called emacqua is the most spoken language in Mozambi.

7:46 It has more than 7 billion speakers. It's of course a band language has like tonal um characteristics. It's a glutenative language and also gender neutral and use like Latin uh script. So in this um map here I'm just showing the the different variant of the language where the most um like the standardized u variant for at least for writing is the the green here in the center of the map. So this um is one that is considered the uh the standard for the for the all the varants that we have for a mark. So the ma is a low resource language which f falls on this um category the category one of taxonomy which means that the language have some label data. However, collecting data is very challenge and this was one of the the difficult part of my research. We didn't have like a lot of data. So, I had to collect and also translate or create like bilingual data for Gumqua and Portuguese.

9:08 And then so I did a lot of data collection and both for training machine translation system also for testing them. So this is me with the translator in Mozambi doing uh this data collection um in Mozambi and from this data collection process I've learned a lot. One of the thing is translating from the higher source language into low source language. It's complex like a in a translation level is very complex and one of the problem is low source language lacks vocabulary coverage. So if we see in slide words like video the visual. So this coming from Portuguese into high resource language into low resource language. So how can you uh translate that into language that doesn't have also covered? So that was uh mostly most of discussion during the data collection or translation process and then one thing that I've learned from there is that the translator usually result on lex code borrowing or

10:27 switching. So let's call boring or long words is something that uh we can see here below in this slide where you borrow a word from the high resource language to into low resource language. So you see words like internet uh it's adapted to fit like the the language as internety something like that. So you adapt that borrowed word into target language and um we have so many examples here. But then the problem is the the adaptation process is not like it's not straightforward and it usually depends on like the preference of the translator. So you could have so different spelling for the same word like that's something that we saw on the existing uh the text data and then the fact that of most of African languages and specifically this language pair that I was working on is morphological complex is native language we need to when you adopting this target language you need to take into account like theization

11:49 rules how to conjugate appropriately also all the adaptation to fits the the the the language the recipient language. So as you can see here in this slide you can see the word radio has been adapted into so many many forms but all of them are arally correct because there's no clear standardization or spelling standards on the language. But but then if you give like a text with one of these spelling option the the person will accept that spelling. But then this is is very well done by humans but it when it comes of machine translation system they kind of uh have difficulty of producing acceptable adaptation. So if you see in this slide the word computer human translation managed to you know adapt to fit like the language rules but the computer and the the empty system here he did he did a poor translation for example the word the word computer he started with the letter C which has a different sound in ma and actually this

13:08 is a violation of the language rule. So if you go to what is happening now uh we do a lot of data synthetic data generation uh with machine translation and imagine that this machine translation system are generating new words that actually not natural. They actually contaminated the the existing index corpora and then like in the long run this is very risk. That's why we uh we ask ourself how can we guide uh machine translation to perform more accurate and appropriate noation in also settings and through the literature we found this uh this concept which is a constraint machine translation which is the concept of you know guiding the generation of the translation output based on particular constraint To give give an example, let's see uh this sentence uh spend the whole night chasing the path. So if you want to translate the sentence, depending on the the domain that you're trying to translate, it could have like a

14:20 different translation. If you're trying to translate like in a biological domain would be like actual b. But in like software engineering is a different thing. So that's the whole point of constraint machine translation where you try to guide the the translation output based on like particular terminology uh of that domain for example. So many approach or strategies exist for doing that like data which we we actually use in this uh particular research and I will explain later. Then nowadays there's a very interesting trend trend which is using LLM following and of course these models are being trained to like set certain checkpoint and then they you can use rag to you know give them more information and this appro this new approach or this new trend to me resembles what we have in the past like the early machine translation system where we have linguists like providing the rules but then they

15:33 didn't have the second part where is the data. So we know that low resource languages uh like realistically talking it will be very difficult to cope with the high resource language because of the limitation in terms of data. But how about we also uh leverage from existing uh information like linguistic information and dictionary grammar etc to also help to produce better translation. That was the my my hypothesis and then just to give more uh explanation on this uh let's suppose that we have a machine translation system that has been trained on the typical way where you have uh feed the system with parallel data and then now you ask uh to translate for example this sentence and then you likely going to have like a wrong output as I explained in my previous slide.

16:29 And then with the the approach that I'm suggesting in this uh uh study is that we can provide some more signals for the more for the model learn how to you know incorporate appropriate um uh translation appropriate uh terminology on the target uh output. So we applied these strong two architectures um which are usually used for machine translation. Uh one is the encoder decoder which is this um first option in which what we did is augmenting the data with this um scheme that we can see here. Uh we have the the the word the potential wrong word and how can you adapt. So we're hoping that uh the model will learn how to copy that uh most possible translation for computer. So we apply this to two multilingual pre-trained models and then M2MP and then we also experiment on the coder only architectures and we we have different models different size but then they all been trained like with the prompt like this one where you

17:50 have the fraction of translate translation plus the guide on how to adapt those those uh long words. So uh for training um the models we use the data set that we have collected. During the process of collecting data, uh we asked the the translator to provide the mappings of of like donor recipient like mango Portuguese when when they are translating if they feel like they result on word adaptation they had to provide that and then we we have those long word annotation. But since our data is not only this uh data set that we collect manually, we also got some datas from the internet but mostly religious uh text and al also other domains. So we also uh weekly annotated this data using uh a model fine tuned from uh GPT4 mini. So what we did was uh finding this model on on the task like this one where we provide source sentence target sentence and the model would output a mappy of donor recipient relationship. So this model

19:15 basically we we train on the mult the data sets that we had the manual annotations and then we evaluated and when we got like 90% of accuracy we got the conf confidence to to use this uh model to annotate the remaining part of the data which didn't have annotation for long models. So this is this was the problem that we use for training the or fine- tuning this model. But then we also evaluate we evaluate on the lores which is data set that we also construct during the data collection process and also has like manual non world annotations.

19:59 And the our idea was uh first evaluate if the machine translation um in terms of quality gets better or det the the rate and for that we use automatic evaluation like blue CHRF and a comment. But we also evaluate using humans, two humans independent uh annotators which basically evaluate each uh translation output on a rate of 0 to 100 and um and then yeah we compute like the agreement between them which was moderate and we also evaluate like what level of accuracy of the output and of the long words and the output generated. So in terms of the training process, we have two setups. The first setup we we train like the typical way of fine-tuning machine translation systems which is just providing parallel data and then we choose to call without constraint and then we have the with constraints which is basically the approach that we are suggesting plus basically we we we had like a mix setup

21:17 where we provide the standard inputs plus we also provide the augmentative input. It's like a bit of redumentation with constraint mechanism and then web was was a way of training by directional uh uh way and for for the LLMs we do some parame parameter efficient methods with Laura and then we explore different sizes of models and this is uh the results that we got. uh so um in the top we have the baseline where as I mentioned that like the typical way of fine-tuning uh using parallel data and then in the bottom we have our method uh well you we can see clearly in all the metrics that we evaluate we got better results uh significant better results and we also noticed that LLMs tend to get more improvements things uh as compared to like baseline. They just better doing the task of adapting onwards. And then as expected this and uh the gains are more uh significant where or greater when when when you

22:33 translating from high resource into low resource like from Portuguese into a macro which was uh what we we expected given the fact this is this is the direction that we mostly have those problem with low long adaptations and we also evaluate the capacitors of the model if you just provide the single like the stand in the model uh train don't want constraint we just provide a the typical or standard input and see whether the model will deteriorate in terms of performance and then we didn't find any deteration in fact some some cases even got better than the the baseline so the model was u robust and stable And then with the humans evaluation, you can see that human preferred models that were trained using our our approach. As you can see here by by the blue violet blocks where they they at the top which corresponded the models that ranked higher in the humans evaluation um in humans evaluation. So what what is it

23:54 like what does it mean in practical sense? Uh let's see let's say this is the the sentence that you want to translate from Portuguese into so the this this sentence the input sentence here is do like like from the plane we saw many hyenas and well bors and when we use our baseline system like typical machine translation fine tuning approach. The translation is completely wrong. Uh which says then we saw a lion cup and a rabbit cup which is uh completely different than distance in terms of me and the sers even with the reference which is here. And with uh with our approach we managed to get more closer um uh translation where where we have also proper targeted sentence. But then we also have some minor uh mistransations here. I mean some minor errors like Ayena and and while boards were were translated like were just mixed so it's just doing just code missing um keeping as it is in the source centers. So uh

25:27 but then the the results are even better than the baseline and we keep asking ourselves how can we improve these results especially on this case where we have identive words that have have been um wrongly translated. So we expand we we ask ourselves this question what that's dictionary over long constraint and then uh from there what we did was just expanding our prompt that we had in the beginning by adding the dictionary information as you can see here and this dictionary information basically we we got it from three uh two three steps process where we first collect different like manual like physical dictionary. We scan them and then we we find a way of creating having each independent entry of dictionary using uh and then like keeping them as JSON file. Then for each sentence pair on the training and test set we did matching or like finding the words in the the sentence that matches with the dictionary entries. So

26:54 to do that we use fuzzy matching and we just keep sentences over or I mean matching that score over 80 85 in terms of matching score and then we basically uh for each matching or the dictionary and long words that we have we construct a prompt just the prompt on something like this and then our model

27:26 and be even a bit a bit of uh better especially for L model and this basically shows that we can improve that those models um in terms of how accurate on on predicting the the the worst or the outputs. So in practical sense uh giving back going back to that example after training the model with that approach we have an even better results here you can see that uh the translation is like from from the plane we saw hyena and wild because the translation is not 100% correct but then because there is a missing word which is uh many. So it doesn't say is he so many but it says he saw aas and and and wild boards. So compared to to the baseline also it's completely uh like this this this this approach provides more close closer um outputs to the reference. So in like an actual production scenario what we envision is something like this. Instead of you having like translation system where you have only

28:59 one box to input your your sentence and you get your translation you could have a option to you know decide how to guide some technology that you have on that uh input sentence. So, so this comes to the conclusion where basically if I have to summarize what this paper is is uh we investigate the impact of finding uh empty systems with donors and dictionary constraint in settings like we focus on the amount of Portuguese language and then we show that this method basically end up enhancing translation quality which is reflected by the different evaluation metrics that we have. And we have evidence that this is um equally positive in multilingual machine translation models or or LLMs.

30:02 And thank you everyone. Happy to get your your questions. I have a question. Yeah, Steve.

30:20 Yeah, please go ahead. Yeah. Uh, thank you Mr. Ad. This is an amazing presentation. Uh, very wonderful and very insightful. Uh, my questions are two. One, More language is a low resource language, right? Uh now considering the models that uh you've experimented was the mark pre-train those models if yes then why do we have to uh integrate or support the training process with uh the long words and then the dictionary that is number one. Number two, uh in your experiments, uh did you ever look at uh the computational resources needed because uh when you do the basic uh translation training using the sentence PS corus straightforward but here you are bringing in additional uh sort of uh additional set of data set to complement the training process. Does that in any way impact on the competitional requirement during the training or compared to where you do the strength forward kind of training without those

31:43 constraint? So those are the two questions. This question is on the whe train on the market. So none of those model have been portrayed on. All right. So uh for the second question uh okay I have I'm sharing here like the output of research but I have tried different things and and in terms of computational

32:16 requirements requirements this process is way less computational require because you only do like pre-processing uh stuff like just just doing the the data documentation at all centers without having to uh impact in the architectural part or you have to like overhead or something like that. And I actually did that which were more computational requiring but uh the results seems to be even better this way. Yeah. Okay. Thank you.

32:57 Thank you. Um, Millison, you have a question in the chat. Do you want to unmute and ask it? Okay, she's asked me to ask a question. So, she says, "Excellent presentation, Felerino. Your method integrates loan word mappings for a MACUA. How well can this approach generalize to other low resource African languages with different borrowing patterns or morphology?"

33:24 That's a good question. uh uh in fact now I'm trying to expand this um like a featur as official fe future work I'm trying to expand this method to other languages um I I presume that my hypothesis that yes this can be expanded to other languages and and uh yeah but I haven't actually did an experiment that's what what I have for like future works that

34:05 Germa I can see your hand raised please go ahead. Yeah, thanks Steph. U Milissant already asked my question on uh scalability to as a low resource Africa language but um Fino great presentation. Thanks. Um before that um my question was like one how how expensive is it to you know when you mention this constraintdriven training and inference if I want to do it for a new language um what is that you know additional uh resource I need or computing I need to include that having said that um what's like the baseline with you know those models out there.

34:52 How is the gap between low resource language and higher ones um narrowing down uh in machine translation? Uh do you see uh we need to have such approach going forward with the latest technologies um out there? So I haven't uh the second question I you you can come after to explain better. I don't think I really understand well but uh the first for the first question in terms of like how expensive is is the is this approach I feel like this is one of the less expensive because what what what was the assumption my research is what how can we use like existing resources like dictionaries which are more available on these low resource languages um than like parallel data Um so this is one but then uh for long long words I feel like this is the most expensive part of of the this resource because uh you need to uh go and you know collect those uh long words mappings between uh source and target sentence.

36:18 But the the there was this approach that I basically used which we can leverage to as a starter right so you can basically use this L&M like GPT and train on on like a prompt like this one where I provide a source sentence target sentence and in this case I had I I just have like um less than thousand examples and then the model m managed to generalize and then I used that model to uh uh basically annotate the remaining part of the model that of the data set that didn't have like manual uh long words annotations. So yeah, I feel like this is one of the less expensive way of of doing the documentation and plus you are bringing the linguistic parts like the the the actual rules of the language to to to to help in generating good translations.

37:27 If I can uh respond the second uh question and you you correct me if I didn't understand well. Um I feel like we still there's still a huge gap between high resource language and low resource language and and uh one of the problem is that to cope with low high resource language depends so heavily having data which is uh almost unrealistic that we going to have in like very short time short short period. So, but then with the with the LLMs now, I feel like this those this LM era are helping to you know getting better models and I I have seen that some of the models are kind of getting acceptable uh translations. Uh yeah, recently yes, I don't know if that was the question.

38:26 Uh it it was uh it was thanks for me know um because I was wondering you know with this new technologies I could see you use JP4 mini as a weak annotation so every time we have a challenge in low resource setting always how can we close that data gap is a key challenge but I feel like it's getting better and better this is an evidence from your slides to use those as weak labels when you have enough loan words and end all that. But um you you correctly addressed my question Santi.

39:09 Great. Um if anyone else has a question, feel free to just unmute yourselves and ask Felerino.

39:15 Yes, go ahead. Uh yes. uh I quite understand the principle behind your training stuff. So you're doing uh an argumentation right just to ensure you have a robust model. So part of those argumentation were basically these long words that were mapped Portuguese and then the mouth work and then the other one the dictionary also. Now the dictionary part of it that as part of was argumentation were they manually annotated or you I think you there is a place where you pinpointing logic or fusing algorithm

39:59 yeah so how did you do that comparison between the ground through dictionaries and then uh how how is that algorithm that is uh uh number one question uh number two question uh I always have multiple questions that's good.

40:14 So number two question now is that uh agumentation using some uh uh manually or existing uh data set to argument with what you have. Uh have you thought of doing the other way around like the back translation? Have you done the comparison where you allow the model to do the translation back and then you turn both the translated and the back original data set and see how the outcome is. So those are the two questions.

40:42 Thank you. Very good questions. Um so the first question in terms of how I like the the fuzzy fuzzy M.

40:54 Yes. Uh so if we see this slide um what like in terms of implementation I didn't want to go like in deeper because that would be like a great presentation. dictionary. How do you do that mapping and get this code to Portuguese?

41:15 So you have the word so you have different words on that source sentence. So we only have the Portuguese toqua diction. So that's why even the slides from here to the end is just Portuguese side and you have each of the word on this sentence. So you compare or not only the word even u you go from like unig grams one words up to like five grams. Yes. And then we compare that with the uh if they match with the the dictionary letters but then the comp the comparison in the dictionary entries is based on on lexical similarity.

42:02 Yeah. So based on the seclex similar similarity which is basically what we have on that fuzzy matching threshold. So if the threshold is higher then of the similarity is higher than 85.

42:17 Yeah. So we we we say this this might be a good match but some but what I'm saying yeah is that some of these matches could be wrong like could be like a a noisy noisy data. Yeah. But um we did some experiment experiment on how this noise data influence on on the quality of translation and that's why we came up with the threshold of H5. So below that you mostly got some noise and then at the bottom. So that's that's what we did and for back translation actually we we came up with this idea after me translation and from my exper experience I didn't really like the bad translation.

43:16 Yeah. So I feel like the back translation was um um was was was risky like creating noisy data as well. Yeah. But the because these systems were not trained on a map. I was creating very very noisy data. That's why I decided to ground the argumentation on like linguistic and actually linguistic information externally information that the translation given the fact that the model were not translation. Yeah, definitely.

44:06 Any additional questions from our remote joiners? Great. Um, Felamino, if you could um, yeah, share any relevant links uh, to your papers in the chat, that would be great. um or any yeah anywhere we can find some of your work online. Um and if that is all I think we can all get 10 minutes back. Thank you everybody for joining. Um we will be posting the recording of this talk on our website in the coming week. Have a good afternoon.

44:48 Thank you. Thank you. Thank you everyone. You cannot take your mother and sleep well.

Questions about this episode

Who speaks in "Leveraging Loanword Constraints for Improving Machine Translation in Low-resource Settings"?

Stephan (Host), Felerino Ali (Guest), Steve (Host), Millison (Guest), Germa (Guest).

What is "Leveraging Loanword Constraints for Improving Machine Translation in Low-resource Settings" about?

The seminar focuses on leveraging loanword constraints to enhance machine translation systems for low-resource languages, with a specific emphasis on the Emakua language. Presenter Felerino Ali outlines methods to incorporate bilingual dictionaries and loanword mappings into training processes, addressing challenges like data scarcity and vocabulary gaps in neural machine translation.

What is mentioned in this Microsoft Research episode?

8 people, works and ideas are mentioned, including Felerino Ali, Emacua, NLP, LLM, Mozambi, Microsoft Research Africa. Each one on this page links to its own page with every episode that mentions it.

How long is "Leveraging Loanword Constraints for Improving Machine Translation in Low-resource Settings"?

The episode runs 45 min; the transcript is about 5,611 words in 50 timestamped sections.

More from Microsoft Research

• All analyzed episodes
• People mentioned on Microsoft Research
• Topics mentioned on Microsoft Research