Fact graph construction process

[ChoiPi2]

Hello! I have a question for you.
First, I built Freebase according to the ToG guidelines and ran LMP main.py to output the results. However, when I look at my output, the graph is not formed for some queries. I don't think the top identity MID of the queries in the freebase exists, but I wonder if the outputs released by LMP were tested with different versions of freebase or are there other issues.

It is being outputted as below for more queries than expected.
ex) "graph": {"Switzerland": {}}}

Also, it seems that the graph is formed by outputting the aggregation result instead of the summary for each relation. For the uploaded outputs, some datasets seem to be composed of the aggregation results, but I wonder if it is correct that the summary results are originally output for each relation.
ex) ../original/lmp_simpleqa_llama-3_1hop.jsonl(aggregation?)
=> {"music.album.primary_release": "The Emissaries has relation music.album.primary_release with following: Emissaries." ~~..
ex) ../original/lmp_webqsp_llama-3_1hop.jsonl(summary)
=> {"people.person.places_lived": "Picasso lived in France, Madrid, and M\u00e1laga.", ~~..

thanks.

[wanjunhong0]

For the first question, we used the same database. If the graph is empty, it means the entity has no tail relation (very rare) or after # of --max_retry the relation sampling and transformation still failed. You and I got different results because the --temperature is set to 0 initially, but will increase 0.2 per retry, where is at line 66 of utils.py. You can run the query with empty graph again and see the retry warning and may have valid output this time.

For the second question, if I only using 1-hop information, I return off and tranformation to save computation cost and there is actually no need to summarize. See line 117 of layers.py. Moreover, in my experience, when I using the old gpt3.5-turbo, sometime it will return the same information of the aggregation and I have to add specific prompt to told GPT do not repeat (this one is removed from current version).

Hope this helps and let me know if you have any other questions.