#TextToSpeech is sounding so advanced now: check out what https://beta.elevenlabs.io can do! #TTS #AI #NaturalLanguage #NaturalLanguageGeneration #NaturalLanguageProcessing #NLP #NeuralNetworks #ArtificialIntelligence #VoiceSynthesis #noxp

#TextToSpeech is sounding so advanced now: check out what https://beta.elevenlabs.io can do! #TTS #AI #NaturalLanguage #NaturalLanguageGeneration #NaturalLanguageProcessing #NLP #NeuralNetworks #ArtificialIntelligence #VoiceSynthesis

Similarly, you may want a so-called #MeaningRepresentation, some structured representation of the information conveyed by a text or a linguistically motivated semantic representation of the text. These annotations are essential for my work in #NaturalLanguageGeneration / #NLG, and a big struggle for the community is coming up with ways to build #corpora for different tasks, domains, genres, or languages which also have the kinds of MRs our systems use.

Yesterday I began writing about #WhatHashtagsMeanToMe, a series of threads where I talk about what the topics highlighted in my profile mean to me.

Today it's about #corpus building, which is the process of finding, building, and/or preparing a linguistic dataset. This is essential work for #NaturalLanguageProcessing and a lot of #AI / #ML systems, as your models can only be as good as the data they are based on.

(Yesterday's thread about #NaturalLanguageGeneration: https://mastodon.social/@_dmh/109303098210409632)

In my own work, I'm working to adapt simpler models which are not as data-hungry to perform data-to-text #NLG in #LowResource settings. Part of this effort will involve exploring #MultitaskLearning and #Pipeline approaches to neural #NaturalLanguageGeneration.

Follow me if you want more content related to these topics, though be advised that this is a whole-person account and I will talk about things other than work as well.

This is not to say, however, that I think these models are useless. I think the interesting question is how to integrate these models into systems that express a particular meaning, a la data-to-text #NaturalLanguageGeneration. Whether this involves #PromptEngineering, integrating them into the decoder for #seq2seq models, or some other more clever application remains to be seen. I am looking forward to seeing how #LLM/s get used for #NLG going forward.

In the years since then there has been an explosive growth in interest in #TextGeneration, but usually not grounded in expressing a particular meaning, which was historically a priority for work in #NaturalLanguageGeneration. This excitement came from the relative fluency of #LargeLanguageModels (#LLM) and #Transformer models in particular, which did an impressive job of continuing an initial utterance.

Around 2015 and 2016 we saw sequence-to-sequence (#seq2seq) models applied to data-to-text #NLG for the first time. These models were trained end-to-end and were very exciting because it raised the prospect of reducing the amount of hand-crafted #GrammarEngineering one would have to do to create a #NaturalLanguageGeneration system.

Since the #NeuralNetwork revolution in #MachineLearning came to #NLP (#NaturalLanguageProcessing/#NLProc), these tools have also become very common in #NaturalLanguageGeneration. Of course, in the late 90s folks were already incorporating statistical information into their #NLG systems in an approach called 'overgenerate-and-rank', where rules were underspecified and produced grammatical and ungrammatical utterances and one relied on n-gram frequencies to rank the possible outputs correctly

#NaturalLanguageGeneration/#NLG has often relied on rule-based approaches to generation with a pipeline of processes. For example, a content selection module might process raw data to decide what is worth mentioning, with the result being fed into a sentence planning module to determine how that information should be expressed broadly (for example, which words to use, how to group the info into sentences, etc), and finally that sentence plan going through 'surface realisation' to become text

But #NaturalLanguageGeneration can broadly apply to all sorts of tasks. Even #MachineTranslation (#MT) is a kind of #NLG. And other text-to-text transformations like summarisation and style-transfer also fall under the umbrella of what I call big-tent #NLG.

For my work, I usually use #NLG to refer to so-called data-to-text generation, where we are transforming some *non-linguistic* representation of information into a natural language utterance (whether that's spoken or written, long or short, formal or informal, etc)

#NaturalLanguageGeneration #data2text #d2t

Every day that I remember to do it, I'm going to explain what I mean by one of the hashtags in my profile. #WhatHashtagsMeanToMe

I'm at the office, so today is about professional stuff: #NaturalLanguageGeneration (#NLG) involves developing computer systems which can express information in natural language (that is, human languages). There are a bunch of different approaches to NLG and a lot of different tasks that can be considered a part of big-tent NLG.

Perhaps time for a new #introduction

I'm a mid-30s he/they kinda guy who works as a researcher in #ComputationalLinguistics (specifically #NLG, #NaturalLanguageGeneration).

I like singing, running, and gaming. I'm queer, polyamorous, Catholic, and an immigrant 3 times over.

Dad to two beautiful voids, my furbabies Salem and Cricket.