Rather than using a massively computational bag-of-words type approach, I was wondering if anyone was approaching the problem this way:
\n\n1) Predict whether the sentence is declarative, imperative, exclamatory or interrogative.
\n2) Then figure out where the parts of speech are.
\n3) See if any are missing
\n4) Use information in 3 to narrow down which word might be missing.
\n\nI don't think that it would negate the machine learning needed to predict the right word, but it might narrow it down.
\nCould people give feedback on the strengths and weakness of looking at the solution this way?
", "messages": [ { "id": "53318", "postDate": "09/08/2014 02:46:16", "content": "Rather than using a massively computational bag-of-words type approach, I was wondering if anyone was approaching the problem this way:
\n\n1) Predict whether the sentence is declarative, imperative, exclamatory or interrogative.
\n2) Then figure out where the parts of speech are.
\n3) See if any are missing
\n4) Use information in 3 to narrow down which word might be missing.
\n\nI don't think that it would negate the machine learning needed to predict the right word, but it might narrow it down.
\nCould people give feedback on the strengths and weakness of looking at the solution this way?
", "rawMarkdown": "", "votes": null }, { "id": "53329", "postDate": "09/08/2014 07:25:15", "content": "It's not clear what you are asking. Predicting the semantic type of a sentence, what all the parts of speech are and guessing what is missing are all machine learning tasks.
\nIf you are asking whether this is a good approach in itself, I'd say that it's not well enough thought out at this stage.
\nIf you are asking whether this would narrow down the search another machine learning approach then you are spending your time in the wrong area. First you've got to build your central machine learning approach, evaluate it on a small amount of the data, then worry about how fast it runs.
\nA first submission should be really simple. Start with a simple idea and make it work. If you start with a whole load of ideas you'll never know what bits are working and what are not so you'll never debug it well enough to get something that works well.
\nIf you want a simple idea to get you started, then I'll write one up.
", "rawMarkdown": "", "votes": null }, { "id": "53335", "postDate": "09/08/2014 10:28:02", "content": "I second wonder if using POS dictionary is against the rules of this competition. The rules say to use training dataset only and POS dictionary would be counted as an additional data.
", "rawMarkdown": "", "votes": null }, { "id": "53347", "postDate": "09/08/2014 14:50:43", "content": "Hi Tony,
\nYes, I should be more clear - predicting the semantic type of a sentence and parts of speech can be machine learning tasks.
\nI was wondering more anyone had a sense of how much value there would be in understanding the structure of a sentence in relation to predicting the missing word. Mainly, I was asking for validation before heading down this path.
\nPer Nicholay's comment - could the organizers clarify if a Parts-of-Speech tagger would be allowed?
", "rawMarkdown": "", "votes": null }, { "id": "53350", "postDate": "09/08/2014 16:14:34", "content": "Traditional Natural Language Processing techniques haven't made a huge difference to the task of Language Modelling. The language modellers started off with simple n-grams and by and large have stayed there. If you have a small amount of training data then class based language models have had some impact, and if you are worried about using a part-of-speech tagger to get these classes then you can generate them automatically (and they work better than parts of speech). That leaves the question as to whether you want to do a full parse or a shallow parse, the problem with full parse is the time it takes in the stochastic framework (i.e. getting a distribution over all parse trees). My tuppence worth is that natural NLP has first to make a difference in speech recognition and machine translation accuracies as published in international conferences and then it's time to read those papers and apply them to this problem.
", "rawMarkdown": "", "votes": null }, { "id": "53352", "postDate": "09/08/2014 17:24:40", "content": "Tony, thanks for your input.
", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 53329, "author_name": "tonyrobinson", "author_url": "", "post_date": "09/08/2014 07:25:15", "content": "It's not clear what you are asking. Predicting the semantic type of a sentence, what all the parts of speech are and guessing what is missing are all machine learning tasks.
\nIf you are asking whether this is a good approach in itself, I'd say that it's not well enough thought out at this stage.
\nIf you are asking whether this would narrow down the search another machine learning approach then you are spending your time in the wrong area. First you've got to build your central machine learning approach, evaluate it on a small amount of the data, then worry about how fast it runs.
\nA first submission should be really simple. Start with a simple idea and make it work. If you start with a whole load of ideas you'll never know what bits are working and what are not so you'll never debug it well enough to get something that works well.
\nIf you want a simple idea to get you started, then I'll write one up.
", "votes": null, "replies": [] }, { "id": 53335, "author_name": "nshmyrev", "author_url": "", "post_date": "09/08/2014 10:28:02", "content": "I second wonder if using POS dictionary is against the rules of this competition. The rules say to use training dataset only and POS dictionary would be counted as an additional data.
", "votes": null, "replies": [] }, { "id": 53347, "author_name": "datawalker", "author_url": "", "post_date": "09/08/2014 14:50:43", "content": "Hi Tony,
\nYes, I should be more clear - predicting the semantic type of a sentence and parts of speech can be machine learning tasks.
\nI was wondering more anyone had a sense of how much value there would be in understanding the structure of a sentence in relation to predicting the missing word. Mainly, I was asking for validation before heading down this path.
\nPer Nicholay's comment - could the organizers clarify if a Parts-of-Speech tagger would be allowed?
", "votes": null, "replies": [] }, { "id": 53350, "author_name": "tonyrobinson", "author_url": "", "post_date": "09/08/2014 16:14:34", "content": "Traditional Natural Language Processing techniques haven't made a huge difference to the task of Language Modelling. The language modellers started off with simple n-grams and by and large have stayed there. If you have a small amount of training data then class based language models have had some impact, and if you are worried about using a part-of-speech tagger to get these classes then you can generate them automatically (and they work better than parts of speech). That leaves the question as to whether you want to do a full parse or a shallow parse, the problem with full parse is the time it takes in the stochastic framework (i.e. getting a distribution over all parse trees). My tuppence worth is that natural NLP has first to make a difference in speech recognition and machine translation accuracies as published in international conferences and then it's time to read those papers and apply them to this problem.
", "votes": null, "replies": [] }, { "id": 53352, "author_name": "datawalker", "author_url": "", "post_date": "09/08/2014 17:24:40", "content": "Tony, thanks for your input.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "53318": "", "53329": "", "53335": "", "53347": "", "53350": "", "53352": "" }, "source": "meta" }